task.py 48.7 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3

4
import os
5
import re
6
import ast
lintangsutawika's avatar
lintangsutawika committed
7
import yaml
lintangsutawika's avatar
lintangsutawika committed
8
import logging
9
10
11
import evaluate
import random
import itertools
12
import functools
13
from tqdm import tqdm
14
15
16
17

import datasets
import numpy as np

baberabb's avatar
baberabb committed
18
from typing import Union, List, Any, Tuple, Literal
19
from collections.abc import Callable
20

21
from lm_eval import utils
22
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
23
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
24
from lm_eval.api.filter import FilterEnsemble
25
26
27

from lm_eval.prompts import get_prompt
from lm_eval.filters import build_filter_ensemble
lintangsutawika's avatar
lintangsutawika committed
28
29
30
31
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
32
    metric_max_over_ground_truths,
lintangsutawika's avatar
lintangsutawika committed
33
34
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
35
36
    get_metric,
    get_aggregation,
37
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
38
    is_higher_better,
39
40
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
41
42
    AGGREGATION_REGISTRY,
)
43

44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
48
    "generate_until",
49
50
]

lintangsutawika's avatar
lintangsutawika committed
51

52
eval_logger = logging.getLogger("lm-eval")
53

lintangsutawika's avatar
lintangsutawika committed
54

55
56
@dataclass
class TaskConfig(dict):
57
    # task naming/registry
58
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
59
    task_alias: str = None
60
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
61
    group_alias: Union[str, list] = None
62
63
64
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
65
66
    dataset_path: str = None
    dataset_name: str = None
67
    dataset_kwargs: dict = None
68
69
70
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
71
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
72
73
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
74
    process_docs: Callable = None
75
76
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
77
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
78
    process_results: Union[Callable, str] = None
79
    use_prompt: str = None
80
    description: str = ""
81
82
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
83
    fewshot_config: dict = None
84
    # runtime configuration options
85
    num_fewshot: int = None
86
    # scoring options
87
    metric_list: list = None
88
    output_type: str = "generate_until"
89
    generation_kwargs: dict = None
90
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
91
    filter_list: Union[str, list] = None
92
93
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
94

lintangsutawika's avatar
lintangsutawika committed
95
96
97
    metadata: Union[
        str, list
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
98

Ethan Smith's avatar
Ethan Smith committed
99
    def __post_init__(self) -> None:
100
        if self.dataset_path and os.path.exists(os.path.dirname(self.dataset_path)):
lintangsutawika's avatar
lintangsutawika committed
101
102
            import inspect
            from importlib import import_module
lintangsutawika's avatar
format  
lintangsutawika committed
103

lintangsutawika's avatar
lintangsutawika committed
104
            self.dataset_path = inspect.getfile(import_module(self.dataset_path))
105

Lintang Sutawika's avatar
Lintang Sutawika committed
106
        if self.generation_kwargs is not None:
107
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
108
                eval_logger.warning(
109
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
110
                )
111
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
112
113
114
115
116
117
118

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

            if "until" not in self.generation_kwargs:
119
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
120
        else:
121
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
122
123
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
124
                    "until": None
125
126
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
127
128
                    "do_sample": False,
                }
129

haileyschoelkopf's avatar
haileyschoelkopf committed
130
131
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

132
133
134
    def __getitem__(self, item):
        return getattr(self, item)

135
136
137
    def __setitem__(self, item, value):
        return setattr(self, item, value)

138
    def to_dict(self):
139
140
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
141
        Used for dumping results alongside full task configuration
142

haileyschoelkopf's avatar
haileyschoelkopf committed
143
144
145
146
147
148
149
150
151
152
        :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
153
154
155
            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
156
        return cfg_dict
157

158
159
160
161
162
163
164
165
166
167
168
169

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
170

171
172
173
174
175
176
177
178
    # 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
179

180
181
182
183
184
185
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
186
    ) -> None:
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
        """
        :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
213
        self._config = TaskConfig({**config}) if config else TaskConfig()
214

lintangsutawika's avatar
lintangsutawika committed
215
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
216

Ethan Smith's avatar
Ethan Smith committed
217
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
        """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.
        """
242
243
244
245
246
247
248
        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,
        )
249

250
251
252
253
254
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
    @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 []

291
292
293
294
295
296
297
298
299
300
    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
301
            eval_logger.warning(
302
                "has_training_docs and has_validation_docs are False"
303
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
304
            )
305
306
            return self.test_docs()

307
308
309
310
311
312
313
314
315
316
    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
317

318
319
320
321
322
323
324
325
326
327
328
329
330
    @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
331
    def doc_to_decontamination_query(self, doc) -> None:
332
333
334
335
336
337
338
339
340
341
342
343
344
        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
345
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
346
347
348
349
350
351
352
353
354
355
        """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!"

356
        eval_logger.info(f"Building contexts for task on rank {rank}...")
357

358
        instances = []
359
360
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
361
        ):
362
            # sample fewshot context #TODO: need to offset doc_id by rank now!
363
            fewshot_ctx = self.fewshot_context(
364
                doc,
365
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
366
            )
367

368
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
369
370
371
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
372
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
373
            )
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398

            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
399
            The number of times each instance in a dataset is inferred on. Defaults to 1,
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
            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
435
436
437
438
439
440
441
442
443
444
    @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))

445
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
446
    def fewshot_context(
447
448
449
450
451
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
452
    ):
453
454
455
456
457
458
459
        """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
460
461
462
463
464
        :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.
465
466
467
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
468
469
470
471
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

472
        description = description if description else ""
473
474

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
475
            labeled_examples = ""
476
        else:
lintangsutawika's avatar
lintangsutawika committed
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
            # 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
501
            )
502
503

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
504
        return description + labeled_examples + example
505
506

    def apply_filters(self):
lintangsutawika's avatar
lintangsutawika committed
507
508
        if hasattr(self, "_filters"):
            for f in self._filters:
lintangsutawika's avatar
lintangsutawika committed
509
                f.apply(self._instances, None)
lintangsutawika's avatar
lintangsutawika committed
510
511
512
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
513

baberabb's avatar
baberabb committed
514
    def dump_config(self) -> dict:
515
        """Returns a dictionary representing the task's config.
516
517
518
519
520

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
521
        # (num_fewshot)
522
        return self.config.to_dict()
523

524
525

class ConfigurableTask(Task):
526
    VERSION = "Yaml"
527
    OUTPUT_TYPE = None
528
    CONFIG = None
529
530
531

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
532
    ) -> None:  # TODO no super() call here
533
        # Get pre-configured attributes
534
        self._config = self.CONFIG
535

536
        # Use new configurations if there was no preconfiguration
537
        if self.config is None:
538
            self._config = TaskConfig(**config)
539
540
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
541
            if config is not None:
542
                self._config.__dict__.update(config)
543

544
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
545
546
547
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
548

549
550
551
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
552

553
554
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
555

556
557
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
558

559
560
561
562
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
563

564
        if self.config.metric_list is None:
565
            # TODO: handle this in TaskConfig.__post_init__ ?
566
567
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

568
            for metric_name in _metric_list:
569
570
                metric = get_metric(metric_name)
                self._metric_fn_list[metric_name] = metric
lintangsutawika's avatar
lintangsutawika committed
571
                self._metric_fn_kwargs[metric_name] = {}
572
573
574
575
576
577
578
                self._aggregation_list[metric_name] = metric.aggregation
                # try:
                #     self._aggregation_list[metric_name] = metric.aggregation
                # except:
                #     self._aggregation_list[metric_name] = get_metric_aggregation(
                #         metric_name
                #     )
haileyschoelkopf's avatar
haileyschoelkopf committed
579
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
580
        else:
581
            for metric_config in self.config.metric_list:
582
583
584
585
586
                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
587
588
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
589
                }
Chris's avatar
Chris committed
590
591
592
593
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
594

595
                if self.config.process_results is not None:
596
597
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
598
599
600
601
602
603
                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
604
605
606
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
607
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
608

609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
                # if "aggregation" in metric_config:
                #     agg_name = metric_config["aggregation"]
                #     if type(agg_name) == str:
                #         self._aggregation_list[metric_name] = get_aggregation(agg_name)
                #     elif callable(agg_name):
                #         self._aggregation_list[metric_name] = metric_config[
                #             "aggregation"
                #         ]
                # else:
                #     INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
                #     metric_agg = get_metric_aggregation(metric_name)
                #     eval_logger.warning(
                #         f"[Task: {self._config.task}] metric {metric_name} is defined, but aggregation is not. "
                #         f"using default "
                #         f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
                #     )
                #     self._aggregation_list[metric_name] = metric_agg

                # if "higher_is_better" in metric_config:
                #     self._higher_is_better[metric_name] = metric_config[
                #         "higher_is_better"
                #     ]
                # else:
                #     eval_logger.warning(
                #         f"[Task: {self._config.task}] metric {metric_name} is defined, but higher_is_better is not. "
                #         f"using default "
                #         f"higher_is_better={is_higher_better(metric_name)}"
                #     )
                #     self._higher_is_better[metric_name] = is_higher_better(metric_name)
638

639
        self.download(self.config.dataset_kwargs)
640
641
642
        self._training_docs = None
        self._fewshot_docs = None

643
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
644
            self._filters = []
645
            for filter_config in self.config.filter_list:
lintangsutawika's avatar
lintangsutawika committed
646
647
648
649
650
651
652
                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
653
654
655
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
656
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
657
        else:
658
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
659

660
661
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
662
            self.prompt = get_prompt(
663
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
664
            )
665
666
667
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
668
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
669
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
670
671
672
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
673
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
674

675
        if self.has_test_docs():
676
            self.task_docs = self.test_docs()
677
        elif self.has_validation_docs():
678
            self.task_docs = self.validation_docs()
679
680
681
682
683
        else:
            assert (
                False
            ), f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"

684
        # Test One Doc
685
        self.features = list(self.task_docs.features.keys())
686
687
        self.multiple_input = 0
        self.multiple_target = 0
688
        test_doc = self.task_docs[0]
689
        test_text = self.doc_to_text(test_doc)
690
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
691

692
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
693
694
695
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
696
697
            else:
                num_choice = len(test_choice)
698

699
700
            if type(test_text) is int:
                self.multiple_input = num_choice
701
702
        else:
            test_choice = None
703

704
        if type(test_target) is list:
705
            self.multiple_target = len(test_target)
706
        else:
lintangsutawika's avatar
lintangsutawika committed
707
            if (type(test_target) is int) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
708
                test_target = test_choice[test_target]
709
            else:
lintangsutawika's avatar
lintangsutawika committed
710
                test_target = str(test_target)
711

712
713
714
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
715
            check_choices = [test_target]
716
717
718
719
        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 = (
720
721
                    True
                    if self.config.target_delimiter.rstrip()
722
                    != self.config.target_delimiter
723
                    else False
724
                )
725

726
727
728
729
730
731
                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(
732
                        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'
733
734
                    )

Ethan Smith's avatar
Ethan Smith committed
735
    def download(self, dataset_kwargs=None) -> None:
736
737
738
739
740
741
        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
742
    def has_training_docs(self) -> bool:
743
        if self.config.training_split is not None:
744
745
746
747
            return True
        else:
            return False

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

baberabb's avatar
baberabb committed
754
    def has_test_docs(self) -> bool:
755
        if self.config.test_split is not None:
756
757
758
759
            return True
        else:
            return False

baberabb's avatar
baberabb committed
760
    def training_docs(self) -> datasets.Dataset:
761
        if self.has_training_docs():
762
763
764
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
765
                )
766
            return self.dataset[self.config.training_split]
767

baberabb's avatar
baberabb committed
768
    def validation_docs(self) -> datasets.Dataset:
769
        if self.has_validation_docs():
770
771
772
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
773
                )
774
            return self.dataset[self.config.validation_split]
775

baberabb's avatar
baberabb committed
776
    def test_docs(self) -> datasets.Dataset:
777
        if self.has_test_docs():
778
779
780
            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]
781

782
    def fewshot_docs(self):
783
784
        if self.config.fewshot_split is not None:
            return self.dataset[self.config.fewshot_split]
785
        else:
786
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
787
                eval_logger.warning(
788
                    f"Task '{self.config.task}': "
789
790
791
792
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
793

lintangsutawika's avatar
lintangsutawika committed
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
    @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)

827
828
829
830
831
832
833
834
    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

835
    def should_decontaminate(self):
836
        return self.config.should_decontaminate
837
838

    def doc_to_decontamination_query(self, doc):
839
        if self.config.should_decontaminate:
840
841
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
842
            else:
843
844
845
846
847
848
849
850
851
852
853
                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
                        )
                    )
854

855
856
857
858
859
860
861
862
863
864
865
866
    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):
867
868
        if self.prompt is not None:
            doc_to_text = self.prompt
869
        else:
870
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
871

872
873
874
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
875
            if doc_to_text in self.features:
876
                # if self.config.doc_to_choice is not None:
877
878
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
879
880
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
881
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
882
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
883
884
885
                    return ast.literal_eval(text_string)
                else:
                    return text_string
886
        elif callable(doc_to_text):
887
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
888
        # Used when applying a Promptsource template
889
        elif hasattr(doc_to_text, "apply"):
890
891
892
893
894
            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")
895
                return self.config.fewshot_delimiter
896
        else:
897
            print(type(doc_to_text))
898
            raise TypeError
899

900
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
901
902
        if self.prompt is not None:
            doc_to_target = self.prompt
903
        else:
904
            doc_to_target = self.config.doc_to_target
905

906
907
908
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
909
            if doc_to_target in self.features:
910
                # if self.config.doc_to_choice is not None:
911
912
913
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
914
            else:
lintangsutawika's avatar
lintangsutawika committed
915
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
916
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
917
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
918
919
920
921
922
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
923
924
925
926
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
927
928
                else:
                    return target_string
929
930
        elif type(doc_to_target) == list:
            return doc_to_target
931
        elif callable(doc_to_target):
932
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
933
        # Used when applying a Promptsource template
934
        elif hasattr(doc_to_target, "apply"):
935
            applied_prompt = doc_to_target.apply(doc)
936
937
938
939
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
940
                return self.config.fewshot_delimiter
941
942
        else:
            raise TypeError
943

baberabb's avatar
baberabb committed
944
    def doc_to_choice(self, doc: Any) -> List[str]:
945
946
        if self.prompt is not None:
            doc_to_choice = self.prompt
947
        elif self.config.doc_to_choice is None:
948
949
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
950
            doc_to_choice = self.config.doc_to_choice
951
952

        if type(doc_to_choice) == str:
953
954
955
956
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
957
958
959
960
961
962
963
964
965
966
        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
967

baberabb's avatar
baberabb committed
968
969
970
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
971
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
972
            arguments = (ctx, self.doc_to_target(doc))
973
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
974
            arguments = (self.doc_to_target(doc),)
975
        elif self.OUTPUT_TYPE == "multiple_choice":
976
            choices = self.doc_to_choice(doc)
977
            target_delimiter = self.config.target_delimiter
978
979
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
980
                cont = self.doc_to_target(doc)
981
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
982
            else:
983
                # Otherwise they are placed in the continuation
984
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
985

986
            request_list = [
987
988
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
989
                    doc=doc,
990
                    arguments=arg,
991
                    idx=i,
992
993
                    **kwargs,
                )
994
                for i, arg in enumerate(arguments)
995
            ]
996
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
997
            if "acc_mutual_info" in self._metric_fn_list.keys():
998
999
1000
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1001
                # here mutual info refers to calculating
1002
1003
1004
1005
1006
1007
                # 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
1008
                            doc=doc,
1009
                            arguments=("", "{}".format(choice)),
1010
1011
1012
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1013
                        for i, choice in enumerate(choices)
1014
1015
1016
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1017

1018
        elif self.OUTPUT_TYPE == "generate_until":
1019
            arguments = (ctx, self.config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
1020
1021

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1022
1023
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1024
1025

    def process_results(self, doc, results):
1026
1027
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1028

1029
        result_dict = {}
1030
1031
1032
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1033
            return ll, is_greedy
1034
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1035
            (loglikelihood,) = results
1036
1037
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
1038
            return loglikelihood, _words, _bytes
1039
        elif self.OUTPUT_TYPE == "multiple_choice":
1040
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1041

1042
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1043
            choices = self.doc_to_choice(doc)
1044
1045
            completion_len = np.array([float(len(i)) for i in choices])

1046
1047
            if (
                2 * len(choices) == len(lls)
1048
                and "acc_mutual_info" in self._metric_fn_list.keys()
1049
1050
1051
1052
1053
1054
1055
            ):
                # 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]
1056

1057
1058
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1059

1060
1061
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1062
            else:
1063
                gold = self.doc_to_target(doc)
1064
1065
1066

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

Lintang Sutawika's avatar
Lintang Sutawika committed
1076
                if gold == -100:
1077
1078
1079
1080
                    gold_index_error = True

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

1085
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1086
1087
                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
1088
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1089
1090
1091
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1092
                # 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
1093
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1094
1095

            result_dict = {
1096
                **({"acc": acc} if "acc" in use_metric else {}),
1097
1098
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1099
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1100
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1101
1102
            }

1103
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1104
1105
1106
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1107
1108
1109
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1110
        elif self.OUTPUT_TYPE == "generate_until":
1111
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1112
            result = results[0]
1113
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1114
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1115
                # it assumes that doc_to_target returns a number.
1116
1117
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1118
1119
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1120
                gold = list(gold)
Chris's avatar
Chris committed
1121
1122
1123
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1124

lintangsutawika's avatar
lintangsutawika committed
1125
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1126
1127
1128
1129
1130
                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
1131
1132
1133
1134
                    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
1135
                    for gold_option in gold:
1136
                        try:
1137
                            result_score = self._metric_fn_list[metric](
1138
1139
                                references=[gold_option],
                                predictions=[result],
1140
                                **self._metric_fn_kwargs[metric],
1141
                            )
baberabb's avatar
baberabb committed
1142
1143
1144
                        except (
                            TypeError
                        ):  # TODO: this is hacky and I don't want to do it
1145
                            result_score = self._metric_fn_list[metric](
haileyschoelkopf's avatar
haileyschoelkopf committed
1146
1147
1148
                                [gold_option, result]
                            )
                        if isinstance(result_score, dict):
haileyschoelkopf's avatar
haileyschoelkopf committed
1149
                            # TODO: this handles the case where HF evaluate returns a dict.
1150
                            result_score = result_score[metric]
haileyschoelkopf's avatar
haileyschoelkopf committed
1151
                        scores.append(result_score)
haileyschoelkopf's avatar
haileyschoelkopf committed
1152
                    if any(scores):
1153
                        result_score = 1.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1154
                    else:
1155
                        result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1156
                else:
1157
                    try:
1158
                        result_score = self._metric_fn_list[metric](
1159
1160
                            references=[gold],
                            predictions=[result],
1161
                            **self._metric_fn_kwargs[metric],
1162
                        )
baberabb's avatar
baberabb committed
1163
1164
1165
                    except (
                        TypeError
                    ):  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1166
                        result_score = self._metric_fn_list[metric]([gold, result])
1167
1168
1169
1170
                    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
1171
        else:
lintangsutawika's avatar
lintangsutawika committed
1172
1173
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1174
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1175
            )
1176
1177
1178
1179

        return result_dict

    def aggregation(self):
1180
1181
        # return self._aggregation_list
        return self._metric_fn_list
1182
1183

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1184
        return self._higher_is_better
1185
1186
1187
1188
1189


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

baberabb's avatar
baberabb committed
1190
    def doc_to_target(self, doc: dict) -> str:
1191
1192
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1193
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1194
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1195
1196
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1197
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1198
                doc=doc,
1199
                arguments=(ctx, " {}".format(choice)),
1200
                idx=i,
1201
1202
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1203
1204
            for i, choice in enumerate(doc["choices"])
        ]
1205

baberabb's avatar
baberabb committed
1206
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1207
1208
1209
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
        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
1221
    def higher_is_better(self) -> dict:
1222
1223
1224
1225
1226
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1227
    def aggregation(self) -> dict:
1228
1229
1230
1231
1232
1233
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1234
class PerplexityTask(Task):
1235
1236
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1237
    def has_training_docs(self) -> bool:
1238
1239
        return False

baberabb's avatar
baberabb committed
1240
    def fewshot_examples(self, k: int, rnd) -> List:
1241
1242
1243
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1244
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1245
1246
1247
1248
1249
1250
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1251
    def higher_is_better(self) -> dict:
1252
1253
1254
1255
1256
1257
1258
1259
1260
        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
1261
    def doc_to_text(self, doc) -> str:
1262
1263
1264
1265
1266
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1267
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1268
1269
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1270
1271
1272
1273
1274
1275
1276
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1277

baberabb's avatar
baberabb committed
1278
    def process_results(self, doc: dict, results: float) -> dict:
1279
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1280
1281
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1282
1283
1284
1285
1286
1287
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1288
    def aggregation(self) -> dict:
1289
1290
1291
1292
1293
1294
1295
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1296
    def count_bytes(cls, doc) -> int:
1297
1298
1299
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1300
    def count_words(cls, doc) -> int:
1301
1302
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))