task.py 52.5 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
8
from dataclasses import asdict, dataclass
9
from inspect import getsource
10
from typing import Any, List, Literal, Tuple, Union
11
12
13
14
15

import datasets
import numpy as np

from lm_eval import utils
16
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
17
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
18
from lm_eval.api.metrics import (
19
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
20
21
22
23
    mean,
    weighted_perplexity,
)
from lm_eval.api.registry import (
24
25
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
26
    get_aggregation,
27
    get_metric,
28
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
29
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
30
)
31
32
33
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

34

35
36
37
38
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
39
    "generate_until",
40
41
]

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

lintangsutawika's avatar
lintangsutawika committed
44

45
46
@dataclass
class TaskConfig(dict):
47
    # task naming/registry
48
    task: str = None
lintangsutawika's avatar
lintangsutawika committed
49
    task_alias: str = None
50
    group: Union[str, list] = None
lintangsutawika's avatar
lintangsutawika committed
51
    group_alias: Union[str, list] = None
52
53
54
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
55
56
    dataset_path: str = None
    dataset_name: str = None
57
    dataset_kwargs: dict = None
58
59
60
    training_split: str = None
    validation_split: str = None
    test_split: str = None
lintangsutawika's avatar
lintangsutawika committed
61
    fewshot_split: str = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
62
63
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
64
    process_docs: Callable = None
65
66
    doc_to_text: Union[Callable, str] = None
    doc_to_target: Union[Callable, str] = None
lintangsutawika's avatar
lintangsutawika committed
67
    doc_to_choice: Union[Callable, str, dict, list] = None
lintangsutawika's avatar
lintangsutawika committed
68
    process_results: Union[Callable, str] = None
69
    use_prompt: str = None
70
    description: str = ""
71
72
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
haileyschoelkopf's avatar
haileyschoelkopf committed
73
    fewshot_config: dict = None
74
    # runtime configuration options
75
    num_fewshot: int = None
76
    # scoring options
77
    metric_list: list = None
78
79
80
81
82
83
    output_type: Literal[
        "loglikelihood",
        "loglikelihood_rolling",
        "generate_until",
        "multiple_choice",
    ] = "generate_until"
84
    generation_kwargs: dict = None
85
    repeats: int = 1
lintangsutawika's avatar
lintangsutawika committed
86
    filter_list: Union[str, list] = None
87
88
    should_decontaminate: bool = False
    doc_to_decontamination_query: str = None
89
    metadata: dict = 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:
Lintang Sutawika's avatar
Lintang Sutawika committed
92
        if self.generation_kwargs is not None:
93
            if self.output_type != "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
94
                eval_logger.warning(
95
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
96
                )
97
                assert self.output_type != "generate_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
98
99
100
101
102
103
104

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

            if "until" not in self.generation_kwargs:
105
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
106
        else:
107
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
108
109
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
110
                    "until": None
111
112
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
113
114
                    "do_sample": False,
                }
115

116
117
118
    def __getitem__(self, item):
        return getattr(self, item)

119
120
121
    def __setitem__(self, item, value):
        return setattr(self, item, value)

122
    def to_dict(self, keep_callable: bool = False) -> dict:
123
124
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
125
        Used for dumping results alongside full task configuration
126

haileyschoelkopf's avatar
haileyschoelkopf committed
127
128
129
130
131
132
133
134
135
136
        :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)
137
138
139
140
141
142
143
144
145
146
            elif k == "metric_list":
                for metric_dict in v:
                    for metric_key, metric_value in metric_dict.items():
                        if callable(metric_value):
                            metric_dict[metric_key] = self.serialize_function(
                                metric_value, keep_callable=keep_callable
                            )
                cfg_dict[k] = v
            elif callable(v):
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
haileyschoelkopf's avatar
haileyschoelkopf committed
147
        return cfg_dict
148

149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
    def serialize_function(
        self, value: Union[Callable, str], keep_callable=False
    ) -> Union[Callable, str]:
        """Serializes a given function or string.

        If 'keep_callable' is True, the original callable is returned.
        Otherwise, attempts to return the source code of the callable using 'getsource'.
        """
        if keep_callable:
            return value
        else:
            try:
                return getsource(value)
            except (TypeError, OSError):
                return str(value)

165
166
167
168
169
170
171
172
173
174
175
176

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
177

178
179
180
181
182
183
184
185
    # 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
186

187
188
189
190
191
192
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
Ethan Smith's avatar
Ethan Smith committed
193
    ) -> None:
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
        """
        :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
220
        self._config = TaskConfig({**config}) if config else TaskConfig()
221

lintangsutawika's avatar
lintangsutawika committed
222
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
223

Ethan Smith's avatar
Ethan Smith committed
224
    def download(self, data_dir=None, cache_dir=None, download_mode=None) -> None:
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
        """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.
        """
249
250
251
252
253
254
255
        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,
        )
256

257
258
259
260
261
    @property
    def config(self):
        """Returns the TaskConfig associated with this class."""
        return self._config

262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
    @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 []

298
299
300
301
302
303
304
305
306
307
    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
308
            eval_logger.warning(
309
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
310
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
311
            )
312
313
            return self.test_docs()

314
315
316
317
318
319
320
321
322
323
    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
324

325
326
327
328
329
330
331
332
333
334
335
336
337
    @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
338
    def doc_to_decontamination_query(self, doc) -> None:
339
340
341
342
343
344
345
346
347
348
349
350
351
        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
352
    def build_all_requests(self, limit=None, rank=None, world_size=None) -> None:
353
354
355
356
357
358
        """Build a set of Instances for a task, and store them in task.instances"""
        if self.has_test_docs():
            docs = self.test_docs()
        elif self.has_validation_docs():
            docs = self.validation_docs()
        else:
359
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
360

Baber Abbasi's avatar
Baber Abbasi committed
361
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
362

363
        instances = []
364
365
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
366
        ):
367
            # sample fewshot context #TODO: need to offset doc_id by rank now!
368
            fewshot_ctx = self.fewshot_context(
369
                doc,
370
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
371
            )
372

373
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
374
375
376
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
377
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
378
            )
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403

            if not isinstance(inst, list):
                inst = [inst]

            instances.extend(inst)

        self._instances = instances
        assert len(self._instances) != 0, "task.build_requests() did not find any docs!"

    @abc.abstractmethod
    def construct_requests(self, doc, ctx, **kwargs):
        """Uses RequestFactory to construct Requests and returns an iterable of
        Requests which will be sent to the LM.

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
            The context string, generated by fewshot_context. This includes the natural
            language description, as well as the few shot examples, and the question
            part of the document for `doc`.
        :param doc_idx: int
            The index of a document within `self.test_docs()` or `self.validation_docs()`,
            whichever is the main split used.
        :param repeats: int
        TODO: update this docstring
lintangsutawika's avatar
lintangsutawika committed
404
            The number of times each instance in a dataset is inferred on. Defaults to 1,
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
    def process_results(self, doc, results):
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
        the metric for that one document

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
        """
        pass

    @abc.abstractmethod
    def aggregation(self):
        """
        :returns: {str: [metric_score] -> float}
            A dictionary where keys are the names of submetrics and values are
            functions that aggregate a list of metric scores
        """
        pass

    @abc.abstractmethod
    def higher_is_better(self):
        """
        :returns: {str: bool}
            A dictionary where keys are the names of submetrics and values are
            whether a higher value of the submetric is better
        """
        pass

440
441
442
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
443
444
445
446
447
448
449
450
451
452
    @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))

453
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
454
    def fewshot_context(
455
456
457
458
459
        self,
        doc,
        num_fewshot,
        rnd=random.Random(1234),
        description=None,
lintangsutawika's avatar
lintangsutawika committed
460
    ):
461
462
463
464
465
466
467
        """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
468
469
470
471
472
        :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.
473
474
475
        :returns: str
            The fewshot context.
        """
lintangsutawika's avatar
lintangsutawika committed
476
477
478
479
        assert (
            rnd is not None
        ), "A `random.Random` generator argument must be provided to `rnd`"

480
        description = description if description else ""
481
482

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
483
            labeled_examples = ""
484
        else:
lintangsutawika's avatar
lintangsutawika committed
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
            # 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
509
            )
510
511

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
512
        return description + labeled_examples + example
513
514

    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
515
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
516
517
        if hasattr(self, "_filters"):
            for f in self._filters:
518
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
519
520
521
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
522

baberabb's avatar
baberabb committed
523
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
524
        """Returns the config as a dictionary."""
525
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
526
        # (num_fewshot)
527
        return self.config.to_dict()
528

Baber Abbasi's avatar
Baber Abbasi committed
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if key is None:
            raise ValueError("Key must be provided.")

        if update:
            current_value = getattr(self._config, key, {})
            if not isinstance(current_value, dict):
                raise TypeError(
                    f"Expected a dict for key '{key}', got {type(current_value).__name__} instead."
                )
            current_value.update(value)
        else:
            setattr(self._config, key, value)

    def override_metric(self, metric_name: str) -> None:
        """
        Override the default metrics used for evaluation with custom metrics.

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
        (
            self._metric_fn_list,
            self._aggregation_list,
            self._metric_fn_kwargs,
            self._higher_is_better,
        ) = ({}, {}, {}, {})
        self._metric_fn_list[metric_name] = get_metric(metric_name)
        self._aggregation_list[metric_name] = get_metric_aggregation(metric_name)
        self._higher_is_better[metric_name] = is_higher_better(metric_name)
        self._metric_fn_kwargs[metric_name] = {}
        if not isinstance(self, ConfigurableTask):
            self.process_results = lambda x, y: {metric_name: get_metric(metric_name)}
            self.aggregation = lambda: {
                metric_name: get_metric_aggregation(metric_name)
            }
        setattr(self._config, "metric_list", [{"metric": metric_name}])
        setattr(self._config, "process_results", None)

569
570

class ConfigurableTask(Task):
571
    VERSION = "Yaml"
572
    OUTPUT_TYPE = None
573
    CONFIG = None
574
575
576

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
Ethan Smith's avatar
Ethan Smith committed
577
    ) -> None:  # TODO no super() call here
578
        # Get pre-configured attributes
579
        self._config = self.CONFIG
580

581
        # Use new configurations if there was no preconfiguration
582
        if self.config is None:
583
            self._config = TaskConfig(**config)
584
585
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
586
            if config is not None:
587
                self._config.__dict__.update(config)
588

589
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
590
591
592
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
593

594
595
596
597
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

598
599
600
        if self.config.output_type is not None:
            assert self.config.output_type in ALL_OUTPUT_TYPES
            self.OUTPUT_TYPE = self.config.output_type
601

602
603
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
604

605
606
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
607

608
609
610
611
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
612

613
        if self.config.metric_list is None:
614
            # TODO: handle this in TaskConfig.__post_init__ ?
615
616
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

617
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
618
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
619
                self._metric_fn_kwargs[metric_name] = {}
620
621
622
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
623
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
624
        else:
625
            for metric_config in self.config.metric_list:
626
627
628
629
630
                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
631
632
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
633
                }
Chris's avatar
Chris committed
634
635
636
637
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
638

639
                if self.config.process_results is not None:
640
641
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
642
643
644
645
646
647
                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
648
649
650
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
651
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
652

653
                if "aggregation" in metric_config:
654
                    agg_name = metric_config["aggregation"]
655
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
656
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
657
                    elif callable(agg_name):  # noqa: E721
658
659
660
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
661
                else:
662
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
663
                    metric_agg = get_metric_aggregation(metric_name)
664
                    eval_logger.warning(
665
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
666
667
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
668
                    )
669
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
670

671
672
673
674
675
676
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
677
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
678
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
679
                        f"higher_is_better={is_higher_better(metric_name)}"
680
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
681
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
682

683
        self.download(self.config.dataset_kwargs)
684
685
686
        self._training_docs = None
        self._fewshot_docs = None

687
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
688
            self._filters = []
689
            for filter_config in self.config.filter_list:
690
691
692
693
694
695
696
697
698
                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"
                    }
                    components.append([function["function"], kwargs])
                filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
699
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
700
        else:
701
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
702

703
704
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
705
            self.prompt = get_prompt(
706
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
707
            )
708
709
710
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
711
        if self.fewshot_docs() is not None:
haileyschoelkopf's avatar
haileyschoelkopf committed
712
            self.sampler = samplers.get_sampler(
haileyschoelkopf's avatar
haileyschoelkopf committed
713
714
715
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
haileyschoelkopf's avatar
haileyschoelkopf committed
716
            )(list(self.fewshot_docs()), self, rnd=random.Random(1234))
717

718
        if self.has_test_docs():
719
            self.task_docs = self.test_docs()
720
        elif self.has_validation_docs():
721
            self.task_docs = self.validation_docs()
722
        else:
723
            assert False, f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
724

725
        # Test One Doc
726
        self.features = list(self.task_docs.features.keys())
727
728
        self.multiple_input = 0
        self.multiple_target = 0
729
        test_doc = self.task_docs[0]
730
        test_text = self.doc_to_text(test_doc)
731
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
732

733
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
734
            test_choice = self.doc_to_choice(test_doc)
735
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
736
                eval_logger.error("doc_to_choice must return list")
737
738
            else:
                num_choice = len(test_choice)
739

740
            if isinstance(test_text, int):
741
                self.multiple_input = num_choice
742
743
        else:
            test_choice = None
744

745
        if isinstance(test_target, list):
746
            self.multiple_target = len(test_target)
747
        else:
748
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
749
                test_target = test_choice[test_target]
750
            else:
lintangsutawika's avatar
lintangsutawika committed
751
                test_target = str(test_target)
752

753
754
755
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
756
            check_choices = [test_target]
757
758
759
760
        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 = (
761
762
                    True
                    if self.config.target_delimiter.rstrip()
763
                    != self.config.target_delimiter
764
                    else False
765
                )
766

767
                if delimiter_has_whitespace and choice_has_whitespace:
768
769
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
770
771
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
772
                    eval_logger.debug(
773
                        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'
774
775
                    )

Ethan Smith's avatar
Ethan Smith committed
776
    def download(self, dataset_kwargs=None) -> None:
777
778
779
780
781
782
        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
783
    def has_training_docs(self) -> bool:
784
        if self.config.training_split is not None:
785
786
787
788
            return True
        else:
            return False

baberabb's avatar
baberabb committed
789
    def has_validation_docs(self) -> bool:
790
        if self.config.validation_split is not None:
791
792
793
794
            return True
        else:
            return False

baberabb's avatar
baberabb committed
795
    def has_test_docs(self) -> bool:
796
        if self.config.test_split is not None:
797
798
799
800
            return True
        else:
            return False

baberabb's avatar
baberabb committed
801
    def training_docs(self) -> datasets.Dataset:
802
        if self.has_training_docs():
803
804
805
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
806
                )
807
            return self.dataset[self.config.training_split]
808

baberabb's avatar
baberabb committed
809
    def validation_docs(self) -> datasets.Dataset:
810
        if self.has_validation_docs():
811
812
813
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
814
                )
815
            return self.dataset[self.config.validation_split]
816

baberabb's avatar
baberabb committed
817
    def test_docs(self) -> datasets.Dataset:
818
        if self.has_test_docs():
819
820
821
            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]
822

823
    def fewshot_docs(self):
824
        if self.config.fewshot_split is not None:
825
826
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
827
            return self.dataset[self.config.fewshot_split]
828
        else:
829
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
830
                eval_logger.warning(
831
                    f"Task '{self.config.task}': "
832
833
834
835
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
836

lintangsutawika's avatar
lintangsutawika committed
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
    @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)
859
860
861
862
863
864
865
866
867
868
869
870
871
        if self.multiple_input:
            return labeled_examples
        else:
            if isinstance(example, str):
                return labeled_examples + example
            elif isinstance(example, list):
                return [labeled_examples + ex for ex in example]
            elif isinstance(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)
lintangsutawika's avatar
lintangsutawika committed
872

873
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
874
        """Iterates over FilterEnsembles and applies them to instances"""
875
876
        if hasattr(self, "_filters"):
            for f in self._filters:
877
                f.apply(self._instances)
878
879
880
881
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

882
    def should_decontaminate(self):
883
        return self.config.should_decontaminate
884
885

    def doc_to_decontamination_query(self, doc):
886
        if self.config.should_decontaminate:
887
888
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
889
            else:
890
891
892
893
894
895
896
897
898
899
900
                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
                        )
                    )
901

902
903
904
905
906
907
908
909
910
911
912
913
    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):
914
915
        if self.prompt is not None:
            doc_to_text = self.prompt
916
        else:
917
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
918

919
        if isinstance(doc_to_text, int):
920
            return doc_to_text
921
        elif isinstance(doc_to_text, str):
922
            if doc_to_text in self.features:
923
                # if self.config.doc_to_choice is not None:
924
925
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
926
927
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
928
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
929
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
930
931
932
                    return ast.literal_eval(text_string)
                else:
                    return text_string
933
        elif callable(doc_to_text):
934
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
935
        # Used when applying a Promptsource template
936
        elif hasattr(doc_to_text, "apply"):
937
938
939
940
941
            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")
942
                return self.config.fewshot_delimiter
943
        else:
944
            print(type(doc_to_text))
945
            raise TypeError
946

947
    def doc_to_target(self, doc: dict) -> Union[int, str, list]:
948
949
        if self.prompt is not None:
            doc_to_target = self.prompt
950
        else:
951
            doc_to_target = self.config.doc_to_target
952

953
        if isinstance(doc_to_target, int):
954
            return doc_to_target
955
        elif isinstance(doc_to_target, str):
956
            if doc_to_target in self.features:
957
                # if self.config.doc_to_choice is not None:
958
959
960
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
961
            else:
lintangsutawika's avatar
lintangsutawika committed
962
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
963
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
964
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
965
966
967
968
969
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
970
971
972
973
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
974
975
                else:
                    return target_string
976
        elif isinstance(doc_to_target, list):
977
            return doc_to_target
978
        elif callable(doc_to_target):
979
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
980
        # Used when applying a Promptsource template
981
        elif hasattr(doc_to_target, "apply"):
982
            applied_prompt = doc_to_target.apply(doc)
983
984
985
986
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
987
                return self.config.fewshot_delimiter
988
989
        else:
            raise TypeError
990

baberabb's avatar
baberabb committed
991
    def doc_to_choice(self, doc: Any) -> List[str]:
992
993
        if self.prompt is not None:
            doc_to_choice = self.prompt
994
        elif self.config.doc_to_choice is None:
995
996
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
997
            doc_to_choice = self.config.doc_to_choice
998

999
        if isinstance(doc_to_choice, str):
1000
1001
1002
1003
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1004
        elif isinstance(doc_to_choice, list):
1005
            return doc_to_choice
1006
        elif isinstance(doc_to_choice, dict):
1007
1008
1009
1010
1011
1012
1013
            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
1014

baberabb's avatar
baberabb committed
1015
1016
1017
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1018
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1019
            arguments = (ctx, self.doc_to_target(doc))
1020
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1021
            arguments = (self.doc_to_target(doc),)
1022
        elif self.OUTPUT_TYPE == "multiple_choice":
1023
            choices = self.doc_to_choice(doc)
1024
            target_delimiter = self.config.target_delimiter
1025
1026
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1027
                cont = self.doc_to_target(doc)
1028
1029
1030
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1031
            else:
1032
                # Otherwise they are placed in the continuation
1033
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1034

1035
            request_list = [
1036
1037
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1038
                    doc=doc,
1039
                    arguments=arg,
1040
                    idx=i,
1041
1042
                    **kwargs,
                )
1043
                for i, arg in enumerate(arguments)
1044
            ]
1045
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1046
            if "acc_mutual_info" in self._metric_fn_list.keys():
1047
1048
1049
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1050
                # here mutual info refers to calculating
1051
1052
1053
1054
1055
1056
                # 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
1057
                            doc=doc,
1058
                            arguments=("", "{}".format(choice)),
1059
1060
1061
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1062
                        for i, choice in enumerate(choices)
1063
1064
1065
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1066

1067
        elif self.OUTPUT_TYPE == "generate_until":
1068
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1069
1070

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1071
1072
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1073
1074

    def process_results(self, doc, results):
1075
1076
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1077

1078
        result_dict = {}
1079
        use_metric = list(self._metric_fn_list.keys())
1080
1081
1082
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1083
1084
1085
1086
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1087
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1088
            (loglikelihood,) = results
1089
1090
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1091
            return {
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
                **(
                    {"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
1107
            }
1108
        elif self.OUTPUT_TYPE == "multiple_choice":
1109
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1110

1111
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1112
            choices = self.doc_to_choice(doc)
1113
1114
            completion_len = np.array([float(len(i)) for i in choices])

1115
1116
            if (
                2 * len(choices) == len(lls)
1117
                and "acc_mutual_info" in self._metric_fn_list.keys()
1118
1119
1120
1121
1122
1123
1124
            ):
                # 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]
1125

1126
1127
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1128

1129
1130
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1131
            else:
1132
                gold = self.doc_to_target(doc)
1133
1134

            gold_index_error = False
1135
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1136
1137
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1138
1139
                    gold_index_error = True
            else:
1140
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1141
                    gold = gold if gold < len(choices) else -100
1142
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1143
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1144

Lintang Sutawika's avatar
Lintang Sutawika committed
1145
                if gold == -100:
1146
1147
1148
1149
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1150
                    f"Label index was not in within range of available choices,"
1151
1152
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1153

1154
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1155
1156
                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
1157
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1158
1159
1160
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1161
                # 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
1162
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1163
1164

            result_dict = {
1165
                **({"acc": acc} if "acc" in use_metric else {}),
1166
1167
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1168
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1169
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
1170
1171
            }

1172
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1173
1174
1175
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1176
1177
1178
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1179
        elif self.OUTPUT_TYPE == "generate_until":
1180
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1181
            result = results[0]
1182
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1183
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1184
                # it assumes that doc_to_target returns a number.
1185
1186
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1187
1188
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1189
                gold = list(gold)
Chris's avatar
Chris committed
1190
1191
1192
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1193

lintangsutawika's avatar
lintangsutawika committed
1194
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1195
1196
1197
1198
1199
                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
1200
1201
1202
1203
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1204
1205
1206
1207
1208
1209
1210
1211
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1212
                    else:
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1234
                else:
1235
                    try:
1236
                        result_score = self._metric_fn_list[metric](
1237
1238
                            references=[gold],
                            predictions=[result],
1239
                            **self._metric_fn_kwargs[metric],
1240
                        )
1241
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1242
                        result_score = self._metric_fn_list[metric]([gold, result])
1243
1244
1245
1246
                    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
1247
        else:
lintangsutawika's avatar
lintangsutawika committed
1248
1249
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1250
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1251
            )
1252
1253
1254

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1255
    def aggregation(self) -> dict:
1256
1257
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1258
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1259
        return self._higher_is_better
1260

Baber Abbasi's avatar
Baber Abbasi committed
1261
1262
1263
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1264
1265
1266
1267

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

baberabb's avatar
baberabb committed
1268
    def doc_to_target(self, doc: dict) -> str:
1269
1270
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1271
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1272
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1273
1274
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1275
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1276
                doc=doc,
1277
                arguments=(ctx, " {}".format(choice)),
1278
                idx=i,
1279
1280
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1281
1282
            for i, choice in enumerate(doc["choices"])
        ]
1283

baberabb's avatar
baberabb committed
1284
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1285
1286
1287
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
        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
1299
    def higher_is_better(self) -> dict:
1300
1301
1302
1303
1304
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1305
    def aggregation(self) -> dict:
1306
1307
1308
1309
1310
1311
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1312
class PerplexityTask(Task):
1313
1314
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1315
    def has_training_docs(self) -> bool:
1316
1317
        return False

baberabb's avatar
baberabb committed
1318
    def fewshot_examples(self, k: int, rnd) -> List:
1319
1320
1321
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1322
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1323
1324
1325
1326
1327
1328
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1329
    def higher_is_better(self) -> dict:
1330
1331
1332
1333
1334
1335
1336
1337
1338
        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
1339
    def doc_to_text(self, doc) -> str:
1340
1341
1342
1343
1344
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1345
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1346
1347
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1348
1349
1350
1351
1352
1353
1354
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1355

baberabb's avatar
baberabb committed
1356
    def process_results(self, doc: dict, results: float) -> dict:
1357
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1358
1359
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1360
1361
1362
1363
1364
1365
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1366
    def aggregation(self) -> dict:
1367
1368
1369
1370
1371
1372
1373
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1374
    def count_bytes(cls, doc) -> int:
1375
1376
1377
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1378
    def count_words(cls, doc) -> int:
1379
1380
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))