task.py 42.1 KB
Newer Older
1
import abc
2
from dataclasses import dataclass, field, asdict
3
4

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

import datasets
import numpy as np

baberabb's avatar
baberabb committed
16
from typing import Union, List, Any, Tuple, Literal
17
from collections.abc import Callable
18

19
from lm_eval import utils
20
from lm_eval.api import samplers
haileyschoelkopf's avatar
haileyschoelkopf committed
21
from lm_eval.api.instance import Instance
lintangsutawika's avatar
lintangsutawika committed
22
from lm_eval.api.filter import FilterEnsemble
23
24
25
26

from lm_eval.logger import eval_logger
from lm_eval.prompts import get_prompt
from lm_eval.filters import build_filter_ensemble
lintangsutawika's avatar
lintangsutawika committed
27
28
29
30
from lm_eval.api.metrics import (
    mean,
    weighted_perplexity,
    bits_per_byte,
lintangsutawika's avatar
lintangsutawika committed
31
    metric_max_over_ground_truths,
lintangsutawika's avatar
lintangsutawika committed
32
33
)
from lm_eval.api.registry import (
haileyschoelkopf's avatar
haileyschoelkopf committed
34
35
36
37
    get_metric,
    get_aggregation,
    get_default_aggregation,
    is_higher_better,
38
39
    DEFAULT_METRIC_REGISTRY,
    OUTPUT_TYPE_REGISTRY,
lintangsutawika's avatar
lintangsutawika committed
40
41
    AGGREGATION_REGISTRY,
)
42

43
44
45
46
47
48
49
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
    "greedy_until",
]

50
51
52

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

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

91
    def __post_init__(self):
92

Lintang Sutawika's avatar
Lintang Sutawika committed
93
94
95
        if self.generation_kwargs is not None:
            if self.output_type != "greedy_until":
                eval_logger.warning(
96
                    "passed `generation_kwargs`, but not using `output_type: greedy_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
97
                )
98
                assert self.output_type != "greedy_until"
Lintang Sutawika's avatar
Lintang Sutawika committed
99
100
101
102
103
104
105

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

            if "until" not in self.generation_kwargs:
106
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
107
108
109
110
        else:
            if self.output_type == "greedy_until":
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
Lintang Sutawika's avatar
Lintang Sutawika committed
111
                    "until": None
112
113
                    if self.fewshot_delimiter is None
                    else [self.fewshot_delimiter],
Lintang Sutawika's avatar
Lintang Sutawika committed
114
115
116
                    "do_sample": False,
                    "temperature": 0.0,
                }
117

haileyschoelkopf's avatar
haileyschoelkopf committed
118
119
        # TODO: how to make TaskConfigs be de- and re-serializable, even when using the !function constructor?

120
121
122
    def __getitem__(self, item):
        return getattr(self, item)

123
124
125
    def __setitem__(self, item, value):
        return setattr(self, item, value)

126
    def to_dict(self):
127
128
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
129
        Used for dumping results alongside full task configuration
130

haileyschoelkopf's avatar
haileyschoelkopf committed
131
132
133
134
135
136
137
138
139
140
        :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
141
142
143
            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
144
        return cfg_dict
145

146
147
148
149
150
151
152
153
154
155
156
157

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
158

159
160
161
162
163
164
165
166
    # 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
167

168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
    def __init__(
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config=None,
    ):
        """
        :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.
        """
        self.download(data_dir, cache_dir, download_mode)
        self._training_docs = None
        self._fewshot_docs = None
        self._instances = None

haileyschoelkopf's avatar
haileyschoelkopf committed
202
        self._config = TaskConfig(**config) if config else TaskConfig()
203
204
205

        if not hasattr(self, "_filters"):
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
206
            for name, components in self._config.get(
207
                "filters", [["none", [["take_first", None]]]]
lintangsutawika's avatar
lintangsutawika committed
208
            ):
209
210
211
                filter_pipeline = build_filter_ensemble(name, components)
                self._filters.append(filter_pipeline)

lintangsutawika's avatar
lintangsutawika committed
212
        self.sampler = samplers.Sampler(
213
214
            list(self.fewshot_docs()), self, rnd=random.Random(1234)
        )
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240

    def download(self, data_dir=None, cache_dir=None, download_mode=None):
        """Downloads and returns the task dataset.
        Override this method to download the dataset from a custom API.

        :param data_dir: str
            Stores the path to a local folder containing the `Task`'s data files.
            Use this to specify the path to manually downloaded data (usually when
            the dataset is not publicly accessible).
        :param cache_dir: str
            The directory to read/write the `Task` dataset. This follows the
            HuggingFace `datasets` API with the default cache directory located at:
                `~/.cache/huggingface/datasets`
            NOTE: You can change the cache location globally for a given process
            by setting the shell environment variable, `HF_DATASETS_CACHE`,
            to another directory:
                `export HF_DATASETS_CACHE="/path/to/another/directory"`
        :param download_mode: datasets.DownloadMode
            How to treat pre-existing `Task` downloads and data.
            - `datasets.DownloadMode.REUSE_DATASET_IF_EXISTS`
                Reuse download and reuse dataset.
            - `datasets.DownloadMode.REUSE_CACHE_IF_EXISTS`
                Reuse download with fresh dataset.
            - `datasets.DownloadMode.FORCE_REDOWNLOAD`
                Fresh download and fresh dataset.
        """
241
242
243
244
245
246
247
        self.dataset = datasets.load_dataset(
            path=self.DATASET_PATH,
            name=self.DATASET_NAME,
            data_dir=data_dir,
            cache_dir=cache_dir,
            download_mode=download_mode,
        )
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284

    @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 []

285
286
287
288
289
290
291
292
293
294
    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
295
            eval_logger.warning(
296
                "has_training_docs and has_validation_docs are False"
297
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
298
            )
299
300
            return self.test_docs()

301
302
303
304
305
306
307
308
309
310
    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
311

312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
    @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)

    def doc_to_decontamination_query(self, doc):
        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

339
    def build_all_requests(self, limit=None, rank=None, world_size=None):
340
341
342
343
344
345
346
347
348
349
        """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!"

350
351
352
353
        eval_logger.info(
            f"Building contexts for task '{self._config.task}' on rank {rank}..."
        )

354
        instances = []
355
356
        for doc_id, doc in utils.create_iterator(
            enumerate(docs), rank, world_size, limit
lintangsutawika's avatar
lintangsutawika committed
357
        ):
358
            # sample fewshot context #TODO: need to offset doc_id by rank now!
359
            fewshot_ctx = self.fewshot_context(
360
361
                doc,
                self._config.num_fewshot,
362
            )
363

haileyschoelkopf's avatar
haileyschoelkopf committed
364
            # TODO: we should override self._config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
365
366
367
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
368
                metadata=(self._config["task"], doc_id, self._config.repeats),
lintangsutawika's avatar
lintangsutawika committed
369
            )
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394

            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
395
            The number of times each instance in a dataset is inferred on. Defaults to 1,
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
            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
431
432
433
434
435
436
437
438
439
440
    @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))

441
    @utils.positional_deprecated
442
    def fewshot_context(self, doc, num_fewshot):
443
444
445
446
447
448
449
450
451
452
453
454
        """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:
455
456
            # always prepend the (possibly empty) task description
            labeled_examples = self._config.description
457
        else:
lintangsutawika's avatar
lintangsutawika committed
458
459
460
            labeled_examples = self._config.description + self.sampler.get_context(
                doc, num_fewshot
            )
461
462

        example = self.doc_to_text(doc)
463
464
465
466
        if type(example) == str:
            return labeled_examples + example
        elif type(example) == list:
            return [labeled_examples + ex for ex in example]
467
468
469
        elif type(example) == int:
            choices = self.doc_to_choice(doc)
            return labeled_examples + choices[example]
470
471
472

    def apply_filters(self):

lintangsutawika's avatar
lintangsutawika committed
473
474
475
476
477
478
        if hasattr(self, "_filters"):
            for f in self._filters:
                f.apply(self._instances)
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
479

baberabb's avatar
baberabb committed
480
    def dump_config(self) -> dict:
481
        """Returns a dictionary representing the task's config.
482
483
484
485
486

        :returns: str
            The fewshot context.
        """
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
487
        # (num_fewshot)
488
489
        return self._config.to_dict()

490
491

class ConfigurableTask(Task):
492
    VERSION = "Yaml"
493
    OUTPUT_TYPE = None
494
    CONFIG = None
495
496
497

    def __init__(
        self, data_dir=None, cache_dir=None, download_mode=None, config: dict = None
baberabb's avatar
baberabb committed
498
    ):  # TODO no super() call here
499
        # Get pre-configured attributes
500
        self._config = self.CONFIG
501

502
503
        # Use new configurations if there was no preconfiguration
        if self._config is None:
504
            self._config = TaskConfig(**config)
505
506
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
507
            if config is not None:
508
                self._config.__dict__.update(config)
509

510
        if self._config is None:
lintangsutawika's avatar
lintangsutawika committed
511
512
513
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
514
515

        if self._config.output_type is not None:
516
            assert self._config.output_type in ALL_OUTPUT_TYPES
517
518
            self.OUTPUT_TYPE = self._config.output_type

519
520
521
522
523
524
        if self._config.dataset_path is not None:
            self.DATASET_PATH = self._config.dataset_path

        if self._config.dataset_name is not None:
            self.DATASET_NAME = self._config.dataset_name

525
526
527
528
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
529

530
        _metric_list = DEFAULT_METRIC_REGISTRY[self._config.output_type]
531
        if self._config.metric_list is None:
532
            # TODO: handle this in TaskConfig.__post_init__ ?
533
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
534
535
                self._metric_fn_list[metric_name] = get_metric(metric_name)
                self._aggregation_list[metric_name] = get_default_aggregation(
536
                    metric_name
haileyschoelkopf's avatar
haileyschoelkopf committed
537
538
                )
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
539
540
541
542
543
544
545
546
547
        else:
            for metric_config in self._config.metric_list:
                assert "metric" in metric_config
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key not in ["metric", "aggregation", "higher_is_better"]
                }
548

549
                if self._config.process_results is not None:
550
551
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
552
553
554
555
556
557
558
559
                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:
                    self._metric_fn_list[metric_name] = get_metric(metric_name)
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
560

561
                if "aggregation" in metric_config:
562
                    agg_name = metric_config["aggregation"]
563
                    if type(agg_name) == str:
haileyschoelkopf's avatar
haileyschoelkopf committed
564
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
565
566
567
568
                    elif callable(agg_name):
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
569
                else:
570
571

                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
haileyschoelkopf's avatar
haileyschoelkopf committed
572
                    metric_agg = get_default_aggregation(metric_name)
573
                    eval_logger.warning(
574
575
576
                        f"metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
577
                    )
578
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
579

580
581
582
583
584
585
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
586
587
                        f"metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
588
                        f"higher_is_better={is_higher_better(metric_name)}"
589
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
590
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
591

592
        self.download(self._config.dataset_kwargs)
593
594
595
        self._training_docs = None
        self._fewshot_docs = None

lintangsutawika's avatar
lintangsutawika committed
596
        if self._config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
597
            self._filters = []
lintangsutawika's avatar
lintangsutawika committed
598
599
600
601
602
603
604
605
            for filter_config in self._config.filter_list:
                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
606
607
608
                        }
                        components.append([function["function"], kwargs])
                    filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
609
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
610
        else:
611
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
612
613

        if self._config.use_prompt is not None:
lintangsutawika's avatar
lintangsutawika committed
614
            eval_logger.info(f"loading prompt {self._config.use_prompt}")
615
            self.prompt = get_prompt(
lintangsutawika's avatar
lintangsutawika committed
616
617
                self._config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
            )
618
619
620
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
621
622
        if self.fewshot_docs() is not None:
            self.sampler = samplers.Sampler(
623
                list(self.fewshot_docs()), self, rnd=random.Random(1234)
624
            )
625

626
627
628
629
630
631
632
633
634
        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!"

635
        # Test One Doc
636
637
638
        self.features = list(docs.features.keys())
        self.multiple_input = 0
        self.multiple_target = 0
639
640
        test_doc = docs[0]
        test_text = self.doc_to_text(test_doc)
641
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
642
643
644
645
646

        if self._config.doc_to_choice is not None:
            test_choice = self.doc_to_choice(test_doc)
            if type(test_choice) is not list:
                eval_logger.error("doc_to_choice must return list")
647
648
            else:
                num_choice = len(test_choice)
649

650
651
            if type(test_text) is int:
                self.multiple_input = num_choice
652

653
        if type(test_target) is list:
654
655
            self.multiple_target = len(test_target)

656
657
658
659
660
661
662
663
    def download(self, dataset_kwargs=None):

        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
664
    def has_training_docs(self) -> bool:
665
666
667
668
669
        if self._config.training_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
670
    def has_validation_docs(self) -> bool:
671
672
673
674
675
        if self._config.validation_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
676
    def has_test_docs(self) -> bool:
677
678
679
680
681
        if self._config.test_split is not None:
            return True
        else:
            return False

baberabb's avatar
baberabb committed
682
    def training_docs(self) -> datasets.Dataset:
683
        if self.has_training_docs():
684
            if self._config.process_docs is not None:
685
686
687
                return self._config.process_docs(
                    self.dataset[self._config.training_split]
                )
688
689
            return self.dataset[self._config.training_split]

baberabb's avatar
baberabb committed
690
    def validation_docs(self) -> datasets.Dataset:
691
        if self.has_validation_docs():
692
            if self._config.process_docs is not None:
693
694
695
                return self._config.process_docs(
                    self.dataset[self._config.validation_split]
                )
696
697
            return self.dataset[self._config.validation_split]

baberabb's avatar
baberabb committed
698
    def test_docs(self) -> datasets.Dataset:
699
        if self.has_test_docs():
700
            if self._config.process_docs is not None:
701
                return self._config.process_docs(self.dataset[self._config.test_split])
702
703
            return self.dataset[self._config.test_split]

704
    def fewshot_docs(self):
705
        if self._config.fewshot_split is not None:
706
            return self.dataset[self._config.fewshot_split]
707
708
709
        else:
            if self._config.num_fewshot > 0:
                eval_logger.warning(
haileyschoelkopf's avatar
haileyschoelkopf committed
710
                    f"Task '{self._config.task}': "
711
712
713
714
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
715

716
717
718
719
720
    def should_decontaminate(self):
        return self._config.should_decontaminate

    def doc_to_decontamination_query(self, doc):
        if self._config.should_decontaminate:
721
722
723
724
725
726
            if self._config.doc_to_decontamination_query in self.features:
                return doc[self._config.doc_to_decontamination_query]
            else:
                return ast.literal_eval(
                    utils.apply_template(self._config.doc_to_decontamination_query, doc)
                )
727

728
729
730
731
732
733
734
735
736
737
738
739
    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):
740
741
742

        if self.prompt is not None:
            doc_to_text = self.prompt
743
744
        else:
            doc_to_text = self._config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
745

746
747
748
        if type(doc_to_text) == int:
            return doc_to_text
        elif type(doc_to_text) == str:
749
            if doc_to_text in self.features:
750
751
752
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
753
754
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
755
756
757
758
759
                text_string = utils.apply_template(doc_to_text, doc)
                if text_string.isdigit():
                    return ast.literal_eval(text_string)
                else:
                    return text_string
760
        elif callable(doc_to_text):
761
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
762
        # Used when applying a Promptsource template
763
        elif hasattr(doc_to_text, "apply"):
764
765
766
767
768
            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")
769
                return applied_prompt[0]
770
        else:
771
            print(type(doc_to_text))
772
            raise TypeError
773

baberabb's avatar
baberabb committed
774
    def doc_to_target(self, doc: dict) -> Union[int, str]:
775
776
777

        if self.prompt is not None:
            doc_to_target = self.prompt
778
779
780
        else:
            doc_to_target = self._config.doc_to_target

781
782
783
        if type(doc_to_target) == int:
            return doc_to_target
        elif type(doc_to_target) == str:
784
            if doc_to_target in self.features:
785
786
787
788
                # if self._config.doc_to_choice is not None:
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
789
            else:
lintangsutawika's avatar
lintangsutawika committed
790
791
792
793
794
                target_string = utils.apply_template(doc_to_target, doc)
                if target_string.isdigit():
                    return ast.literal_eval(target_string)
                else:
                    return target_string
795
        elif callable(doc_to_target):
796
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
797
        # Used when applying a Promptsource template
798
        elif hasattr(doc_to_target, "apply"):
799
            applied_prompt = doc_to_target.apply(doc)
800
801
802
803
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
804
                return applied_prompt[0]
805
806
        else:
            raise TypeError
807

baberabb's avatar
baberabb committed
808
    def doc_to_choice(self, doc: Any) -> List[str]:
809
810
811

        if self.prompt is not None:
            doc_to_choice = self.prompt
lintangsutawika's avatar
lintangsutawika committed
812
        elif self._config.doc_to_choice is None:
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
            doc_to_choice = self._config.doc_to_choice

        if type(doc_to_choice) == str:
            return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
        elif type(doc_to_choice) == list:
            return doc_to_choice
        elif type(doc_to_choice) == dict:
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
        elif hasattr(doc_to_choice, "get_answer_choices_list"):
            return doc_to_choice.get_answer_choices_list(doc)
        else:
            raise TypeError
829

830
    def gold_alias(self, doc):
831
832
833
834
835
        # returns a version of the gold target answer to a document,
        # which should be passed into metric for scoring as the ground truth.

        # in multiple_choice tasks, this should be castable to an int corresponding to the index
        # within the answer choices, while doc_to_target is the string version of {{answer_choices[gold]}}.
lintangsutawika's avatar
lintangsutawika committed
836
        if self._config.gold_alias is not None:
837
838
            doc_to_target = self._config.gold_alias
        else:
lintangsutawika's avatar
lintangsutawika committed
839
            return self.doc_to_target(doc)
840
841
842
843
844
845
846
847
848
849

        if type(doc_to_target) == str:
            return utils.apply_template(doc_to_target, doc)
        elif callable(doc_to_target):
            return doc_to_target(doc)
        elif hasattr(doc_to_target, "apply"):
            return doc_to_target.apply(doc)[1]
        else:
            raise TypeError

baberabb's avatar
baberabb committed
850
851
852
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
853

854
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
855
            arguments = (ctx, self.doc_to_target(doc))
856
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
857
            arguments = (self.doc_to_target(doc),)
858
        elif self.OUTPUT_TYPE == "multiple_choice":
859
860

            choices = self.doc_to_choice(doc)
861
            target_delimiter = self._config.target_delimiter
862
863
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
864
                cont = self.doc_to_target(doc)
865
                arguments = [(ctx, f"{target_delimiter}{cont}") for ctx in choices]
866
            else:
867
                # Otherwise they are placed in the continuation
868
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
869

870
            request_list = [
871
872
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
873
                    doc=doc,
874
                    arguments=arg,
875
                    idx=i,
876
877
                    **kwargs,
                )
878
                for i, arg in enumerate(arguments)
879
            ]
880
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
881
            if "acc_mutual_info" in self._metric_fn_list.keys():
882
883
884
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
885
                # here mutual info refers to calculating
886
887
888
889
890
891
                # 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
892
                            doc=doc,
893
                            arguments=("", "{}".format(choice)),
894
895
896
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
897
                        for i, choice in enumerate(choices)
898
899
900
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
901

902
        elif self.OUTPUT_TYPE == "greedy_until":
903
            arguments = (ctx, self._config.generation_kwargs)
lintangsutawika's avatar
lintangsutawika committed
904
905

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
906
907
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
908
909
910

    def process_results(self, doc, results):

lintangsutawika's avatar
lintangsutawika committed
911
912
        if callable(self._config.process_results):
            return self._config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
913

914
        result_dict = {}
915
        use_metric = list(self._metric_fn_list.keys())
916
917
918
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
919
920
921
922
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
923
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
924
            (loglikelihood,) = results
925
926
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
927
            return {
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
                **(
                    {"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
943
            }
944
        elif self.OUTPUT_TYPE == "multiple_choice":
945
946

            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
947

948
            # retrieve choices in List[str] form, to compute choice lengths, etc.
949
            choices = self.doc_to_choice(doc)
950
951
            completion_len = np.array([float(len(i)) for i in choices])

952
953
            if (
                2 * len(choices) == len(lls)
954
                and "acc_mutual_info" in self._metric_fn_list.keys()
955
956
957
958
959
960
961
            ):
                # 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]
962

963
964
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
965

966
967
            if self.multiple_input:
                gold = self.doc_to_text(doc)
968
            else:
969
                gold = self.doc_to_target(doc)
970
971
                if type(gold) is str:
                    gold = choices.index(gold)
lintangsutawika's avatar
lintangsutawika committed
972

973
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
974
975
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
976
                exact_match = int(any([is_greedy[i] for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
977
978
979
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
980
981
                # TODO: this gets score of 0 on arc_challenge for pythia-70m. need to test that this works properly
                exact_match = int(is_greedy[gold])
982
983

            result_dict = {
984
                **({"acc": acc} if "acc" in use_metric else {}),
985
986
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
987
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
988
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
989
990
            }

991
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
992
993
994
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
995
996
997
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

998
999
        elif self.OUTPUT_TYPE == "greedy_until":

1000
            gold = self.doc_to_target(doc)
1001
1002
1003
            if type(gold) == int:
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1004

1005
            for key, result in zip(self._metric_fn_list.keys(), results):
haileyschoelkopf's avatar
haileyschoelkopf committed
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
                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 = []
                    for gold_option in gold:
                        res = self._metric_fn_list[key](
                            references=[gold_option],
                            predictions=[result],
                            **self._metric_fn_kwargs[key],
                        )
                        if isinstance(res, dict):
                            # TODO: this handles the case where HF evaluate returns a dict.
                            res = res[key]
                        scores.append(res)
                    if any(scores):
                        result = 1.0
                    else:
                        result = 0.0
                else:
                    result = self._metric_fn_list[key](
                        references=[gold],
                        predictions=[result],
                        **self._metric_fn_kwargs[key],
                    )
1031

haileyschoelkopf's avatar
haileyschoelkopf committed
1032
1033
1034
1035
                if isinstance(result, dict):
                    result_dict.update(result)
                else:
                    result_dict[key] = result
1036
        else:
lintangsutawika's avatar
lintangsutawika committed
1037
1038
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1039
                "'loglikelihood', 'loglikelihood_rolling', 'greedy_until' or 'multiple_choice'",
1040
            )
1041
1042
1043
1044
1045
1046
1047

        return result_dict

    def aggregation(self):
        return self._aggregation_list

    def higher_is_better(self):
haileyschoelkopf's avatar
haileyschoelkopf committed
1048
        return self._higher_is_better
1049
1050
1051
1052
1053


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

baberabb's avatar
baberabb committed
1054
    def doc_to_target(self, doc: dict) -> str:
1055
1056
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1057
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1058
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1059
1060
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1061
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1062
                doc=doc,
1063
                arguments=(ctx, " {}".format(choice)),
1064
                idx=i,
1065
1066
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1067
1068
            for i, choice in enumerate(doc["choices"])
        ]
1069

baberabb's avatar
baberabb committed
1070
    def process_results(self, doc: dict, results: List[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1071
1072
1073
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
        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
1085
    def higher_is_better(self) -> dict:
1086
1087
1088
1089
1090
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1091
    def aggregation(self) -> dict:
1092
1093
1094
1095
1096
1097
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1098
class PerplexityTask(Task):
1099
1100
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1101
    def has_training_docs(self) -> bool:
1102
1103
        return False

baberabb's avatar
baberabb committed
1104
    def fewshot_examples(self, k: int, rnd) -> List:
1105
1106
1107
        assert k == 0
        return []

baberabb's avatar
baberabb committed
1108
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1109
1110
1111
1112
1113
1114
        assert (
            num_fewshot == 0
        ), "The number of fewshot examples must be 0 for perplexity tasks."

        return ""

baberabb's avatar
baberabb committed
1115
    def higher_is_better(self) -> dict:
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
        return {
            "word_perplexity": False,
            "byte_perplexity": False,
            "bits_per_byte": False,
        }

    def doc_to_decontamination_query(self, doc):
        return doc

    def doc_to_text(self, doc):
        return ""

    def doc_to_target(self, doc):
        return doc

baberabb's avatar
baberabb committed
1131
    def construct_requests(self, doc: dict, ctx: Union[str, None], **kwargs):
1132
1133
        assert not ctx

lintangsutawika's avatar
lintangsutawika committed
1134
1135
1136
1137
1138
1139
1140
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1141

baberabb's avatar
baberabb committed
1142
    def process_results(self, doc: dict, results: float) -> dict:
1143
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1144
1145
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1146
1147
1148
1149
1150
1151
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1152
    def aggregation(self) -> dict:
1153
1154
1155
1156
1157
1158
1159
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1160
    def count_bytes(cls, doc) -> int:
1161
1162
1163
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1164
    def count_words(cls, doc) -> int:
1165
1166
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))