task.py 48.9 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import os
5
import random
6
7
8
9
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass
from typing import Any, List, Literal, Tuple, Union
10
11
12
13
14

import datasets
import numpy as np

from lm_eval import utils
15
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
16
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
17
from lm_eval.api.metrics import (
18
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
19
20
21
22
    mean,
    weighted_perplexity,
)
from lm_eval.api.registry import (
23
24
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
25
    get_aggregation,
26
    get_metric,
27
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
28
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
29
)
30
31
32
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

33

34
35
36
37
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
38
    "generate_until",
39
40
]

lintangsutawika's avatar
lintangsutawika committed
41

42
eval_logger = logging.getLogger("lm-eval")
43

lintangsutawika's avatar
lintangsutawika committed
44

45
46
@dataclass
class TaskConfig(dict):
47
    # task naming/registry
48
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
49
    task_alias: str = None
50
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
51
    group_alias: Union[str, list] = None
52
53
54
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
55
56
    dataset_path: str = None
    dataset_name: str = None
57
    dataset_kwargs: dict = None
58
59
60
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
61
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
62
63
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
64
    process_docs: Callable = None
65
66
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
67
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
68
    process_results: Union[Callable, str] = None
69
    use_prompt: str = None
70
    description: str = ""
71
72
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
73
    fewshot_config: dict = None
74
    # runtime configuration options
75
    num_fewshot: int = None
76
    # scoring options
77
    metric_list: list = None
78
    output_type: str = "generate_until"
79
    generation_kwargs: dict = None
80
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
81
    filter_list: Union[str, list] = None
82
83
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
84

lintangsutawika's avatar
lintangsutawika committed
85
86
87
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
88

Ethan Smith's avatar
Ethan Smith committed
89
    def __post_init__(self) -> None:
90
        if self.dataset_path and os.path.exists(os.path.dirname(self.dataset_path)):
lintangsutawika's avatar
lintangsutawika committed
91
92
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
93

lintangsutawika's avatar
lintangsutawika committed
94
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
95

Lintang Sutawika's avatar
Lintang Sutawika committed
96
        if self.generation_kwargs is not None:
97
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
98
                eval_logger.warning(
99
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
100
                )
101
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
102
103
104
105
106
107
108

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

            if "until" not in self.generation_kwargs:
109
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
110
        else:
111
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
112
113
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
114
                    "until": None
115
116
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
                    "do_sample": False,
                }
119

haileyschoelkopf's avatar
haileyschoelkopf committed
120
121
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

122
123
124
    def __getitem__(self, item):
        return getattr(self, item)

125
126
127
    def __setitem__(self, item, value):
        return setattr(self, item, value)

128
    def to_dict(self):
129
130
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
131
        Used for dumping results alongside full task configuration
132

haileyschoelkopf's avatar
haileyschoelkopf committed
133
134
135
136
137
138
139
140
141
142
        :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
143
144
145
            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
146
        return cfg_dict
147

148
149
150
151
152
153
154
155
156
157
158
159

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
160

161
162
163
164
165
166
167
168
    # 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
169

170
171
172
173
174
175
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
176
    ) -> None:
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
        """
        :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
203
        self._config = TaskConfig({**config}) if config else TaskConfig()
204

lintangsutawika's avatar
lintangsutawika committed
205
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
206

Ethan Smith's avatar
Ethan Smith committed
207
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
        """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.
        """
232
233
234
235
236
237
238
        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,
        )
239

240
241
242
243
244
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
    @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 []

281
282
283
284
285
286
287
288
289
290
    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
291
            eval_logger.warning(
292
                "has_training_docs and has_validation_docs are False"
293
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
294
            )
295
296
            return self.test_docs()

297
298
299
300
301
302
303
304
305
306
    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
307

308
309
310
311
312
313
314
315
316
317
318
319
320
    @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
321
    def doc_to_decontamination_query(self, doc) -> None:
322
323
324
325
326
327
328
329
330
331
332
333
334
        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
335
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
336
337
338
339
340
341
        """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:
342
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
343

344
        eval_logger.info(f"Building contexts for task on rank {rank}...")
345

346
        instances = []
347
348
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
349
        ):
350
            # sample fewshot context #TODO: need to offset doc_id by rank now!
351
            fewshot_ctx = self.fewshot_context(
352
                doc,
353
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
354
            )
355

356
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
357
358
359
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
360
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
361
            )
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386

            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
387
            The number of times each instance in a dataset is inferred on. Defaults to 1,
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
            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
423
424
425
426
427
428
429
430
431
432
    @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))

433
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
434
    def fewshot_context(
435
436
437
438
439
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
440
    ):
441
442
443
444
445
446
447
        """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
448
449
450
451
452
        :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.
453
454
455
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
456
457
458
459
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

460
        description = description if description else ""
461
462

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
463
            labeled_examples = ""
464
        else:
lintangsutawika's avatar
lintangsutawika committed
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
            # 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
489
            )
490
491

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
492
        return description + labeled_examples + example
493
494

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
495
496
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
497
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
498
499
500
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
501

baberabb's avatar
baberabb committed
502
    def dump_config(self) -> dict:
503
        """Returns a dictionary representing the task's config.
504
505
506
507
508

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
509
        # (num_fewshot)
510
        return self.config.to_dict()
511

512
513

class ConfigurableTask(Task):
514
    VERSION = "Yaml"
515
    OUTPUT_TYPE = None
516
    CONFIG = None
517
518
519

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
520
    ) -> None:  # TODO no super() call here
521
        # Get pre-configured attributes
522
        self._config = self.CONFIG
523

524
        # Use new configurations if there was no preconfiguration
525
        if self.config is None:
526
            self._config = TaskConfig(**config)
527
528
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
529
            if config is not None:
530
                self._config.__dict__.update(config)
531

532
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
533
534
535
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
536

537
538
539
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
540

541
542
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
543

544
545
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
546

547
548
549
550
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
551

552
        if self.config.metric_list is None:
553
            # TODO: handle this in TaskConfig.__post_init__ ?
554
555
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

556
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
557
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
558
                self._metric_fn_kwargs[metric_name] = {}
559
560
561
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
562
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
563
        else:
564
            for metric_config in self.config.metric_list:
565
566
567
568
569
                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
570
571
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
572
                }
Chris's avatar
Chris committed
573
574
575
576
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
577

578
                if self.config.process_results is not None:
579
580
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
581
582
583
584
585
586
                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
587
588
589
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
590
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
591

592
                if "aggregation" in metric_config:
593
                    agg_name = metric_config["aggregation"]
594
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
595
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
596
                    elif callable(agg_name):  # noqa: E721
597
598
599
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
600
                else:
601
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
602
                    metric_agg = get_metric_aggregation(metric_name)
603
                    eval_logger.warning(
baberabb's avatar
baberabb committed
604
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
605
606
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
607
                    )
608
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
609

610
611
612
613
614
615
                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
616
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
617
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
618
                        f"higher_is_better={is_higher_better(metric_name)}"
619
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
620
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
621

622
        self.download(self.config.dataset_kwargs)
623
624
625
        self._training_docs = None
        self._fewshot_docs = None

626
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
627
            self._filters = []
628
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
629
630
631
632
633
634
635
                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
636
637
638
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
639
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
640
        else:
641
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
642

643
644
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
645
            self.prompt = get_prompt(
646
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
647
            )
648
649
650
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
651
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
652
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
653
654
655
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
656
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
657

658
        if self.has_test_docs():
659
            self.task_docs = self.test_docs()
660
        elif self.has_validation_docs():
661
            self.task_docs = self.validation_docs()
662
        else:
663
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
664

665
        # Test One Doc
666
        self.features = list(self.task_docs.features.keys())
667
668
        self.multiple_input = 0
        self.multiple_target = 0
669
        test_doc = self.task_docs[0]
670
        test_text = self.doc_to_text(test_doc)
671
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
672

673
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
674
            test_choice = self.doc_to_choice(test_doc)
675
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
676
                eval_logger.error("doc_to_choice must return list")
677
678
            else:
                num_choice = len(test_choice)
679

680
            if isinstance(test_text, int):
681
                self.multiple_input = num_choice
682
683
        else:
            test_choice = None
684

685
        if isinstance(test_target, list):
686
            self.multiple_target = len(test_target)
687
        else:
688
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
689
                test_target = test_choice[test_target]
690
            else:
lintangsutawika's avatar
lintangsutawika committed
691
                test_target = str(test_target)
692

693
694
695
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
696
            check_choices = [test_target]
697
698
699
700
        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 = (
701
702
                    True
                    if self.config.target_delimiter.rstrip()
703
                    != self.config.target_delimiter
704
                    else False
705
                )
706

707
708
709
710
711
712
                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(
713
                        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'
714
715
                    )

Ethan Smith's avatar
Ethan Smith committed
716
    def download(self, dataset_kwargs=None) -> None:
717
718
719
720
721
722
        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
723
    def has_training_docs(self) -> bool:
724
        if self.config.training_split is not None:
725
726
727
728
            return True
        else:
            return False

baberabb's avatar
baberabb committed
729
    def has_validation_docs(self) -> bool:
730
        if self.config.validation_split is not None:
731
732
733
734
            return True
        else:
            return False

baberabb's avatar
baberabb committed
735
    def has_test_docs(self) -> bool:
736
        if self.config.test_split is not None:
737
738
739
740
            return True
        else:
            return False

baberabb's avatar
baberabb committed
741
    def training_docs(self) -> datasets.Dataset:
742
        if self.has_training_docs():
743
744
745
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
746
                )
747
            return self.dataset[self.config.training_split]
748

baberabb's avatar
baberabb committed
749
    def validation_docs(self) -> datasets.Dataset:
750
        if self.has_validation_docs():
751
752
753
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
754
                )
755
            return self.dataset[self.config.validation_split]
756

baberabb's avatar
baberabb committed
757
    def test_docs(self) -> datasets.Dataset:
758
        if self.has_test_docs():
759
760
761
            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]
762

763
    def fewshot_docs(self):
764
765
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
766
        else:
767
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
768
                eval_logger.warning(
769
                    f"Task '{self.config.task}': "
770
771
772
773
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
774

lintangsutawika's avatar
lintangsutawika committed
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
    @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)
797
        if isinstance(example, str):
lintangsutawika's avatar
lintangsutawika committed
798
            return labeled_examples + example
799
        elif isinstance(example, list):
lintangsutawika's avatar
lintangsutawika committed
800
            return [labeled_examples + ex for ex in example]
801
        elif isinstance(example, int):
lintangsutawika's avatar
lintangsutawika committed
802
803
804
805
806
807
            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)

808
809
810
811
812
813
814
815
    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

816
    def should_decontaminate(self):
817
        return self.config.should_decontaminate
818
819

    def doc_to_decontamination_query(self, doc):
820
        if self.config.should_decontaminate:
821
822
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
823
            else:
824
825
826
827
828
829
830
831
832
833
834
                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
                        )
                    )
835

836
837
838
839
840
841
842
843
844
845
846
847
    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):
848
849
        if self.prompt is not None:
            doc_to_text = self.prompt
850
        else:
851
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
852

853
        if isinstance(doc_to_text, int):
854
            return doc_to_text
855
        elif isinstance(doc_to_text, str):
856
            if doc_to_text in self.features:
857
                # if self.config.doc_to_choice is not None:
858
859
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
860
861
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
862
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
863
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
864
865
866
                    return ast.literal_eval(text_string)
                else:
                    return text_string
867
        elif callable(doc_to_text):
868
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
869
        # Used when applying a Promptsource template
870
        elif hasattr(doc_to_text, "apply"):
871
872
873
874
875
            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")
876
                return self.config.fewshot_delimiter
877
        else:
878
            print(type(doc_to_text))
879
            raise TypeError
880

881
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
882
883
        if self.prompt is not None:
            doc_to_target = self.prompt
884
        else:
885
            doc_to_target = self.config.doc_to_target
886

887
        if isinstance(doc_to_target, int):
888
            return doc_to_target
889
        elif isinstance(doc_to_target, str):
890
            if doc_to_target in self.features:
891
                # if self.config.doc_to_choice is not None:
892
893
894
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
895
            else:
lintangsutawika's avatar
lintangsutawika committed
896
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
897
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
898
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
899
900
901
902
903
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
904
905
906
907
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
908
909
                else:
                    return target_string
910
        elif isinstance(doc_to_target, list):
911
            return doc_to_target
912
        elif callable(doc_to_target):
913
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
914
        # Used when applying a Promptsource template
915
        elif hasattr(doc_to_target, "apply"):
916
            applied_prompt = doc_to_target.apply(doc)
917
918
919
920
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
921
                return self.config.fewshot_delimiter
922
923
        else:
            raise TypeError
924

baberabb's avatar
baberabb committed
925
    def doc_to_choice(self, doc: Any) -> List[str]:
926
927
        if self.prompt is not None:
            doc_to_choice = self.prompt
928
        elif self.config.doc_to_choice is None:
929
930
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
931
            doc_to_choice = self.config.doc_to_choice
932

933
        if isinstance(doc_to_choice, str):
934
935
936
937
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
938
        elif isinstance(doc_to_choice, list):
939
            return doc_to_choice
940
        elif isinstance(doc_to_choice, dict):
941
942
943
944
945
946
947
            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
948

baberabb's avatar
baberabb committed
949
950
951
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
952
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
953
            arguments = (ctx, self.doc_to_target(doc))
954
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
955
            arguments = (self.doc_to_target(doc),)
956
        elif self.OUTPUT_TYPE == "multiple_choice":
957
            choices = self.doc_to_choice(doc)
958
            target_delimiter = self.config.target_delimiter
959
960
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
961
                cont = self.doc_to_target(doc)
962
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
963
            else:
964
                # Otherwise they are placed in the continuation
965
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
966

967
            request_list = [
968
969
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
970
                    doc=doc,
971
                    arguments=arg,
972
                    idx=i,
973
974
                    **kwargs,
                )
975
                for i, arg in enumerate(arguments)
976
            ]
977
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
978
            if "acc_mutual_info" in self._metric_fn_list.keys():
979
980
981
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
982
                # here mutual info refers to calculating
983
984
985
986
987
988
                # 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
989
                            doc=doc,
990
                            arguments=("", "{}".format(choice)),
991
992
993
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
994
                        for i, choice in enumerate(choices)
995
996
997
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
998

999
        elif self.OUTPUT_TYPE == "generate_until":
1000
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1001
1002

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1003
1004
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1005
1006

    def process_results(self, doc, results):
1007
1008
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1009

1010
        result_dict = {}
1011
        use_metric = list(self._metric_fn_list.keys())
1012
1013
1014
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1015
1016
1017
1018
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1019
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1020
            (loglikelihood,) = results
1021
1022
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1023
            return {
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
                **(
                    {"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
1039
            }
1040
        elif self.OUTPUT_TYPE == "multiple_choice":
1041
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1042

1043
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1044
            choices = self.doc_to_choice(doc)
1045
1046
            completion_len = np.array([float(len(i)) for i in choices])

1047
1048
            if (
                2 * len(choices) == len(lls)
1049
                and "acc_mutual_info" in self._metric_fn_list.keys()
1050
1051
1052
1053
1054
1055
1056
            ):
                # 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]
1057

1058
1059
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1060

1061
1062
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1063
            else:
1064
                gold = self.doc_to_target(doc)
1065
1066

            gold_index_error = False
1067
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1068
1069
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1070
1071
                    gold_index_error = True
            else:
1072
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1073
                    gold = gold if gold < len(choices) else -100
1074
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1075
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1076

Lintang Sutawika's avatar
Lintang Sutawika committed
1077
                if gold == -100:
1078
1079
1080
1081
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1082
                    f"Label index was not in within range of available choices,"
1083
1084
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1085

1086
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1087
1088
                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
1089
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1090
1091
1092
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1093
                # 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
1094
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1095
1096

            result_dict = {
1097
                **({"acc": acc} if "acc" in use_metric else {}),
1098
1099
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1100
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1101
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1102
1103
            }

1104
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1105
1106
1107
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1108
1109
1110
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1111
        elif self.OUTPUT_TYPE == "generate_until":
1112
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1113
            result = results[0]
1114
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1115
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1116
                # it assumes that doc_to_target returns a number.
1117
1118
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1119
1120
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1121
                gold = list(gold)
Chris's avatar
Chris committed
1122
1123
1124
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1125

lintangsutawika's avatar
lintangsutawika committed
1126
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1127
1128
1129
1130
1131
                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
1132
1133
1134
1135
                    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
1136
                    for gold_option in gold:
1137
                        try:
1138
                            result_score = self._metric_fn_list[metric](
1139
1140
                                references=[gold_option],
                                predictions=[result],
1141
                                **self._metric_fn_kwargs[metric],
1142
                            )
baberabb's avatar
baberabb committed
1143
1144
1145
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1146
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1147
1148
1149
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1150
                            # TODO: this handles the case where HF evaluate returns a dict.
1151
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1152
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1153
                    if any(scores):
1154
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1155
                    else:
1156
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1157
                else:
1158
                    try:
1159
                        result_score = self._metric_fn_list[metric](
1160
1161
                            references=[gold],
                            predictions=[result],
1162
                            **self._metric_fn_kwargs[metric],
1163
                        )
1164
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1165
                        result_score = self._metric_fn_list[metric]([gold, result])
1166
1167
1168
1169
                    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
1170
        else:
lintangsutawika's avatar
lintangsutawika committed
1171
1172
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1173
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1174
            )
1175
1176
1177
1178
1179
1180
1181

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1182
        return self._higher_is_better
1183
1184
1185
1186
1187


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

baberabb's avatar
baberabb committed
1188
    def doc_to_target(self, doc: dict) -> str:
1189
1190
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1191
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1192
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1193
1194
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1195
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1196
                doc=doc,
1197
                arguments=(ctx, " {}".format(choice)),
1198
                idx=i,
1199
1200
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1201
1202
            for i, choice in enumerate(doc["choices"])
        ]
1203

baberabb's avatar
baberabb committed
1204
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1205
1206
1207
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
        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
1219
    def higher_is_better(self) -> dict:
1220
1221
1222
1223
1224
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1225
    def aggregation(self) -> dict:
1226
1227
1228
1229
1230
1231
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1232
class PerplexityTask(Task):
1233
1234
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1235
    def has_training_docs(self) -> bool:
1236
1237
        return False

baberabb's avatar
baberabb committed
1238
    def fewshot_examples(self, k: int, rnd) -> List:
1239
1240
1241
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1242
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1243
1244
1245
1246
1247
1248
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1249
    def higher_is_better(self) -> dict:
1250
1251
1252
1253
1254
1255
1256
1257
1258
        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
1259
    def doc_to_text(self, doc) -> str:
1260
1261
1262
1263
1264
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1265
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1266
1267
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1268
1269
1270
1271
1272
1273
1274
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1275

baberabb's avatar
baberabb committed
1276
    def process_results(self, doc: dict, results: float) -> dict:
1277
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1278
1279
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1280
1281
1282
1283
1284
1285
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1286
    def aggregation(self) -> dict:
1287
1288
1289
1290
1291
1292
1293
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1294
    def count_bytes(cls, doc) -> int:
1295
1296
1297
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1298
    def count_words(cls, doc) -> int:
1299
1300
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))