task.py 49.2 KB
Newer Older
1
import abc
2
import ast
Herbie Bradley's avatar
Herbie Bradley committed
3
import itertools
lintangsutawika's avatar
lintangsutawika committed
4
import logging
Herbie Bradley's avatar
Herbie Bradley committed
5
6
7
8
9
import random
import re
from collections.abc import Callable
from dataclasses import asdict, dataclass, field
from typing import Any, List, Literal, Tuple, Union
10
11

import datasets
Herbie Bradley's avatar
Herbie Bradley committed
12
import evaluate
13
import numpy as np
Herbie Bradley's avatar
Herbie Bradley committed
14
15
16
import scipy.special as sp
import yaml
from tqdm import tqdm
17

18
from lm_eval import utils
19
from lm_eval.api import samplers
lintangsutawika's avatar
lintangsutawika committed
20
from lm_eval.api.filter import FilterEnsemble
lintangsutawika's avatar
lintangsutawika committed
21
22
from lm_eval.api.metrics import (
    bits_per_byte,
Herbie Bradley's avatar
Herbie Bradley committed
23
    mean,
lintangsutawika's avatar
lintangsutawika committed
24
    metric_max_over_ground_truths,
Herbie Bradley's avatar
Herbie Bradley committed
25
    weighted_perplexity,
lintangsutawika's avatar
lintangsutawika committed
26
27
)
from lm_eval.api.registry import (
Herbie Bradley's avatar
Herbie Bradley committed
28
29
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
30
    get_aggregation,
Herbie Bradley's avatar
Herbie Bradley committed
31
    get_metric,
32
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
33
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
34
)
Herbie Bradley's avatar
Herbie Bradley committed
35
36
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt
37

38
39
40
41
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
42
    "generate_until",
43
44
]

45

46
eval_logger = logging.getLogger("lm-eval")
47

lintangsutawika's avatar
lintangsutawika committed
48

49
50
@dataclass
class TaskConfig(dict):
51
    # task naming/registry
52
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
53
    task_alias: str = None
54
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
55
    group_alias: Union[str, list] = None
56
57
58
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
59
60
    dataset_path: str = None
    dataset_name: str = None
61
    dataset_kwargs: dict = None
62
63
64
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
65
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
66
67
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
68
    process_docs: Callable = None
69
70
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
71
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
72
    process_results: Union[Callable, str] = None
73
    use_prompt: str = None
74
    description: str = ""
75
76
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
77
    fewshot_config: dict = None
78
    # runtime configuration options
79
    num_fewshot: int = 0
80
    # scoring options
81
    metric_list: list = None
82
    output_type: str = "generate_until"
83
    generation_kwargs: dict = None
84
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
85
    filter_list: Union[str, list] = None
86
87
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
88

lintangsutawika's avatar
lintangsutawika committed
89
    metadata: str = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
90

Ethan Smith's avatar
Ethan Smith committed
91
    def __post_init__(self) -> None:
lintangsutawika's avatar
lintangsutawika committed
92
        if self.dataset_path and ("." in self.dataset_path):
lintangsutawika's avatar
lintangsutawika committed
93
94
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
95

lintangsutawika's avatar
lintangsutawika committed
96
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
97

Lintang Sutawika's avatar
Lintang Sutawika committed
98
        if self.generation_kwargs is not None:
99
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
100
                eval_logger.warning(
101
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
102
                )
103
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
108
109
110

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

            if "until" not in self.generation_kwargs:
111
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
112
        else:
113
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
114
115
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
116
                    "until": None
117
118
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
119
120
                    "do_sample": False,
                }
121

haileyschoelkopf's avatar
haileyschoelkopf committed
122
123
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

124
125
126
    def __getitem__(self, item):
        return getattr(self, item)

127
128
129
    def __setitem__(self, item, value):
        return setattr(self, item, value)

130
    def to_dict(self):
131
132
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
133
        Used for dumping results alongside full task configuration
134

haileyschoelkopf's avatar
haileyschoelkopf committed
135
136
137
138
139
140
141
142
143
144
        :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
145
146
147
            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
148
        return cfg_dict
149

150
151
152
153
154
155
156
157
158
159
160
161

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
162

163
164
165
166
167
168
169
170
    # 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
171

172
173
174
175
176
177
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
178
    ) -> None:
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
        """
        :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
205
        self._config = TaskConfig({**config}) if config else TaskConfig()
206

lintangsutawika's avatar
lintangsutawika committed
207
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
208

Ethan Smith's avatar
Ethan Smith committed
209
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
        """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.
        """
234
235
236
237
238
239
240
        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,
        )
241

242
243
244
245
246
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

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
281
282
    @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 []

283
284
285
286
287
288
289
290
291
292
    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
293
            eval_logger.warning(
294
                "has_training_docs and has_validation_docs are False"
295
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
296
            )
297
298
            return self.test_docs()

299
300
301
302
303
304
305
306
307
308
    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
309

310
311
312
313
314
315
316
317
318
319
320
321
322
    @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
323
    def doc_to_decontamination_query(self, doc) -> None:
324
325
326
327
328
329
330
331
332
333
334
335
336
        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
337
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
338
339
340
341
342
343
344
345
346
347
        """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!"

348
        eval_logger.info(f"Building contexts for task on rank {rank}...")
349

350
        instances = []
351
352
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
353
        ):
354
            # sample fewshot context #TODO: need to offset doc_id by rank now!
355
            fewshot_ctx = self.fewshot_context(
356
                doc,
357
                self.config.num_fewshot,
358
            )
359

360
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
361
362
363
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
364
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
365
            )
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390

            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
391
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
423
424
425
426
            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
427
428
429
430
431
432
433
434
435
436
    @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))

437
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
438
    def fewshot_context(
439
440
441
442
443
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
444
    ):
445
446
447
448
449
450
451
        """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
452
453
454
455
456
        :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.
457
458
459
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
460
461
462
463
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

464
        description = description if description else ""
465
466

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
467
            labeled_examples = ""
468
        else:
lintangsutawika's avatar
lintangsutawika committed
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
            # 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
493
            )
494
495

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
496
        return description + labeled_examples + example
497
498

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
499
500
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
501
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
502
503
504
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
505

baberabb's avatar
baberabb committed
506
    def dump_config(self) -> dict:
507
        """Returns a dictionary representing the task's config.
508
509
510
511
512

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
513
        # (num_fewshot)
514
        return self.config.to_dict()
515

516
517

class ConfigurableTask(Task):
518
    VERSION = "Yaml"
519
    OUTPUT_TYPE = None
520
    CONFIG = None
521
522
523

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
524
    ) -> None:  # TODO no super() call here
525
        # Get pre-configured attributes
526
        self._config = self.CONFIG
527

528
        # Use new configurations if there was no preconfiguration
529
        if self.config is None:
530
            self._config = TaskConfig(**config)
531
532
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
533
            if config is not None:
534
                self._config.__dict__.update(config)
535

536
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
537
538
539
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
540

541
542
543
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
544

545
546
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
547

548
549
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
550

551
552
553
554
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
555

556
        if self.config.metric_list is None:
557
            # TODO: handle this in TaskConfig.__post_init__ ?
558
559
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

560
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
561
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
562
                self._metric_fn_kwargs[metric_name] = {}
563
                self._aggregation_list[metric_name] = get_metric_aggregation(
564
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
565
566
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
567
        else:
568
            for metric_config in self.config.metric_list:
569
570
571
572
573
                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
574
575
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
576
                }
Chris's avatar
Chris committed
577
578
579
580
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
581

582
                if self.config.process_results is not None:
583
584
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
585
586
587
588
589
590
                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
591
592
593
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
594
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
595

596
                if "aggregation" in metric_config:
597
                    agg_name = metric_config["aggregation"]
598
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
599
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
600
601
602
603
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
604
                else:
605
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
606
                    metric_agg = get_metric_aggregation(metric_name)
607
                    eval_logger.warning(
baberabb's avatar
baberabb committed
608
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
609
610
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
611
                    )
612
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
613

614
615
616
617
618
619
                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
620
                        f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
621
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
622
                        f"higher_is_better={is_higher_better(metric_name)}"
623
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
624
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
625

626
        self.download(self.config.dataset_kwargs)
627
628
629
        self._training_docs = None
        self._fewshot_docs = None

630
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
631
            self._filters = []
632
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
633
634
635
636
637
638
639
                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
640
641
642
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
643
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
644
        else:
645
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
646

647
648
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
649
            self.prompt = get_prompt(
650
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
651
            )
652
653
654
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
655
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
656
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
657
658
659
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
660
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
661

662
        if self.has_test_docs():
663
            self.task_docs = self.test_docs()
664
        elif self.has_validation_docs():
665
            self.task_docs = self.validation_docs()
666
667
668
669
670
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

671
        # Test One Doc
Herbie Bradley's avatar
Herbie Bradley committed
672
673
        # self.features = ["text", "meta"]
        # return None
674
        self.features = list(self.task_docs.features.keys())
675
676
        self.multiple_input = 0
        self.multiple_target = 0
677
        test_doc = self.task_docs[0]
678
        test_text = self.doc_to_text(test_doc)
679
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
680

681
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
682
683
684
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
685
686
            else:
                num_choice = len(test_choice)
687

688
689
            if type(test_text) is int:
                self.multiple_input = num_choice
690
691
        else:
            test_choice = None
692

693
        if type(test_target) is list:
694
            self.multiple_target = len(test_target)
695
        else:
lintangsutawika's avatar
lintangsutawika committed
696
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
697
                test_target = test_choice[test_target]
698
            else:
lintangsutawika's avatar
lintangsutawika committed
699
                test_target = str(test_target)
700

701
702
703
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
704
            check_choices = [test_target]
705
706
707
708
        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 = (
709
710
                    True
                    if self.config.target_delimiter.rstrip()
711
                    != self.config.target_delimiter
712
                    else False
713
                )
714

715
716
717
718
719
720
                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(
721
                        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'
722
723
                    )

Ethan Smith's avatar
Ethan Smith committed
724
    def download(self, dataset_kwargs=None) -> None:
725
726
727
728
729
730
        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
731
    def has_training_docs(self) -> bool:
732
        if self.config.training_split is not None:
733
734
735
736
            return True
        else:
            return False

baberabb's avatar
baberabb committed
737
    def has_validation_docs(self) -> bool:
738
        if self.config.validation_split is not None:
739
740
741
742
            return True
        else:
            return False

baberabb's avatar
baberabb committed
743
    def has_test_docs(self) -> bool:
744
        if self.config.test_split is not None:
745
746
747
748
            return True
        else:
            return False

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

baberabb's avatar
baberabb committed
757
    def validation_docs(self) -> datasets.Dataset:
758
        if self.has_validation_docs():
759
760
761
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
762
                )
763
            return self.dataset[self.config.validation_split]
764

baberabb's avatar
baberabb committed
765
    def test_docs(self) -> datasets.Dataset:
766
        if self.has_test_docs():
767
768
769
            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]
770

771
    def fewshot_docs(self):
772
773
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
774
        else:
775
            if self.config.num_fewshot > 0:
776
                eval_logger.warning(
777
                    f"Task '{self.config.task}': "
778
779
780
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
Herbie Bradley's avatar
Herbie Bradley committed
781
782
783
                return super().fewshot_docs()
            else:
                return None
784

785
    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
786
787
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
    @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)

819
820
821
822
823
824
825
826
    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

827
    def should_decontaminate(self):
828
        return self.config.should_decontaminate
829
830

    def doc_to_decontamination_query(self, doc):
831
832
833
        if self.config.should_decontaminate:
            if self.config.doc_to_decontamination_query in self.features:
                return doc[self.config.doc_to_decontamination_query]
834
835
            else:
                return ast.literal_eval(
836
                    utils.apply_template(self.config.doc_to_decontamination_query, doc)
837
                )
838

839
840
841
842
843
844
845
846
847
848
849
850
    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):
851
852
        if self.prompt is not None:
            doc_to_text = self.prompt
853
        else:
854
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
855

856
857
858
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
859
            if doc_to_text in self.features:
860
                # if self.config.doc_to_choice is not None:
861
862
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
863
864
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
865
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
866
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
867
868
869
                    return ast.literal_eval(text_string)
                else:
                    return text_string
870
        elif callable(doc_to_text):
871
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
872
        # Used when applying a Promptsource template
873
        elif hasattr(doc_to_text, "apply"):
874
875
876
877
878
            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")
879
                return self.config.fewshot_delimiter
880
        else:
881
            print(type(doc_to_text))
882
            raise TypeError
883

884
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
885
886
        if self.prompt is not None:
            doc_to_target = self.prompt
887
        else:
888
            doc_to_target = self.config.doc_to_target
889

890
891
892
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
893
            if doc_to_target in self.features:
894
                # if self.config.doc_to_choice is not None:
895
896
897
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
898
            else:
lintangsutawika's avatar
lintangsutawika committed
899
                target_string = utils.apply_template(doc_to_target, doc)
Herbie Bradley's avatar
Herbie Bradley committed
900
                # return target_string
lintangsutawika's avatar
lintangsutawika committed
901
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
902
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
903
904
905
906
907
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
908
909
910
911
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
912
913
                else:
                    return target_string
914
915
        elif type(doc_to_target) == list:
            return doc_to_target
916
        elif callable(doc_to_target):
917
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
918
        # Used when applying a Promptsource template
919
        elif hasattr(doc_to_target, "apply"):
920
            applied_prompt = doc_to_target.apply(doc)
921
922
923
924
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
925
                return self.config.fewshot_delimiter
926
927
        else:
            raise TypeError
928

baberabb's avatar
baberabb committed
929
    def doc_to_choice(self, doc: Any) -> List[str]:
930
931
        if self.prompt is not None:
            doc_to_choice = self.prompt
932
        elif self.config.doc_to_choice is None:
933
934
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
935
            doc_to_choice = self.config.doc_to_choice
936
937
938
939
940
941
942
943
944
945
946
947
948

        if type(doc_to_choice) == str:
            return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
        elif type(doc_to_choice) == list:
            return doc_to_choice
        elif type(doc_to_choice) == dict:
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
        elif hasattr(doc_to_choice, "get_answer_choices_list"):
            return doc_to_choice.get_answer_choices_list(doc)
        else:
            raise TypeError
949

baberabb's avatar
baberabb committed
950
951
952
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
953
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
954
            arguments = (ctx, self.doc_to_target(doc))
955
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
956
            arguments = (self.doc_to_target(doc),)
957
        elif self.OUTPUT_TYPE == "multiple_choice":
958
            choices = self.doc_to_choice(doc)
959
            target_delimiter = self.config.target_delimiter
960
961
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
962
                cont = self.doc_to_target(doc)
963
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
964
            else:
965
                # Otherwise they are placed in the continuation
966
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
967

968
            request_list = [
969
970
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
971
                    doc=doc,
972
                    arguments=arg,
973
                    idx=i,
974
975
                    **kwargs,
                )
976
                for i, arg in enumerate(arguments)
977
            ]
978
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
979
            if "acc_mutual_info" in self._metric_fn_list.keys():
980
981
982
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
983
                # here mutual info refers to calculating
984
985
986
987
988
989
                # 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
990
                            doc=doc,
991
                            arguments=("", "{}".format(choice)),
992
993
994
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
995
                        for i, choice in enumerate(choices)
996
997
998
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
999

1000
        elif self.OUTPUT_TYPE == "generate_until":
1001
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1002
1003

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1004
1005
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1006
1007

    def process_results(self, doc, results):
1008
1009
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1010

1011
        result_dict = {}
1012
        use_metric = list(self._metric_fn_list.keys())
1013
1014
1015
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1016
1017
1018
1019
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1020
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1021
            (loglikelihood,) = results
1022
1023
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1024
            return {
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
                **(
                    {"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
1040
            }
1041
        elif self.OUTPUT_TYPE == "multiple_choice":
1042
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1043

1044
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1045
            choices = self.doc_to_choice(doc)
1046
1047
            completion_len = np.array([float(len(i)) for i in choices])

1048
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1049
1050
1051
1052
1053
1054
                # 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]
1055

1056
1057
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1058

1059
1060
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1061
            else:
1062
                gold = self.doc_to_target(doc)
1063
1064
1065

            gold_index_error = False
            if type(gold) is list:
Lintang Sutawika's avatar
Lintang Sutawika committed
1066
1067
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1068
1069
1070
                    gold_index_error = True
            else:
                if type(gold) is int:
Lintang Sutawika's avatar
Lintang Sutawika committed
1071
                    gold = gold if gold < len(choices) else -100
1072
                elif type(gold) is str:
Lintang Sutawika's avatar
Lintang Sutawika committed
1073
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1074

Lintang Sutawika's avatar
Lintang Sutawika committed
1075
                if gold == -100:
1076
1077
1078
1079
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1080
                    f"Label index was not in within range of available choices,"
1081
1082
                    f"Sample:\n\n{doc}\n\n"
                )
1083
1084
1085

            if "ece" in use_metric:
                # Convert lls from log-probabilities to normalized probabilities
1086
1087
                norm_probs: list[float] = np.exp(lls - sp.logsumexp(lls)).tolist()
                calib_scores: list[float] = [0.0] * len(choices)
1088
1089
1090
1091
1092
                if isinstance(gold, list):
                    for g in gold:
                        calib_scores[g] = 1.0
                else:
                    calib_scores[gold] = 1.0
1093
                calibration_probs: dict[str, list[float]] = {
1094
1095
1096
1097
                    "probs": norm_probs,
                    "scores": calib_scores,
                }

1098
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1099
1100
                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
1101
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1102
1103
1104
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1105
                # 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
1106
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1107
1108

            result_dict = {
1109
                **({"acc": acc} if "acc" in use_metric else {}),
1110
1111
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1112
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1113
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1114
                **({"ece": calibration_probs} if "ece" 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
                            )
Herbie Bradley's avatar
Herbie Bradley 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
                        )
Herbie Bradley's avatar
Herbie Bradley 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))