task.py 62.3 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
11
12
13
14
15
16
17
18
19
20
21
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
22
23
24

import datasets
import numpy as np
25
from tqdm import tqdm
26
27

from lm_eval import utils
28
from lm_eval.api import samplers
29
30
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
lintangsutawika's avatar
lintangsutawika committed
31
from lm_eval.api.registry import (
32
33
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
34
    get_aggregation,
35
    get_metric,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
38
)
39
from lm_eval.caching.cache import load_from_cache, save_to_cache
40
41
42
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

43

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

51
eval_logger = logging.getLogger("lm-eval")
52

lintangsutawika's avatar
lintangsutawika committed
53

54
55
@dataclass
class TaskConfig(dict):
56
    # task naming/registry
57
58
59
60
    task: Optional[str] = None
    task_alias: Optional[str] = None
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
61
62
63
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
64
65
66
67
68
69
70
71
72
    dataset_path: Optional[str] = None
    dataset_name: Optional[str] = None
    dataset_kwargs: Optional[dict] = None
    training_split: Optional[str] = None
    validation_split: Optional[str] = None
    test_split: Optional[str] = None
    fewshot_split: Optional[
        str
    ] = None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
73
74
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
75
76
77
78
79
80
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
81
    description: str = ""
82
83
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
84
    fewshot_config: Optional[dict] = None
85
    # runtime configuration options
86
    num_fewshot: Optional[int] = None
87
    # scoring options
88
89
90
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
91
    repeats: int = 1
92
    filter_list: Optional[Union[str, list]] = None
93
    should_decontaminate: bool = False
94
95
96
97
    doc_to_decontamination_query: Optional[str] = None
    metadata: Optional[
        dict
    ] = None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
98

Ethan Smith's avatar
Ethan Smith committed
99
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
100
        if self.generation_kwargs is not None:
101
            if self.output_type != "generate_until":
102
                eval_logger.warning(
103
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
104
105
106
107
108
109
110
111
                )

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

            if "until" not in self.generation_kwargs:
112
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
113
        else:
114
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
115
116
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
117
118
119
120
121
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
122
123
                    "do_sample": False,
                }
124

125
126
127
    def __getitem__(self, item):
        return getattr(self, item)

128
129
130
    def __setitem__(self, item, value):
        return setattr(self, item, value)

131
    def to_dict(self, keep_callable: bool = False) -> dict:
132
133
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
134
        Used for dumping results alongside full task configuration
135

haileyschoelkopf's avatar
haileyschoelkopf committed
136
137
138
139
140
141
142
143
144
145
        :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)
146
147
148
149
150
151
152
153
154
155
            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
156
        return cfg_dict
157

158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
    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)

174
175
176
177
178
179
180
181
182
183
184

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)
    """

185
    VERSION: Optional[Union[int, str]] = None
186

187
188
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
189
    DATASET_PATH: Optional[str] = None
190
191

    # The name of a subset within `DATASET_PATH`.
192
    DATASET_NAME: Optional[str] = None
193

194
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
195

196
197
    def __init__(
        self,
198
199
200
201
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode: Optional[datasets.DownloadMode] = None,
        config: Optional[Mapping] = None,  # Union[dict, TaskConfig]
Ethan Smith's avatar
Ethan Smith committed
202
    ) -> None:
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
        """
        :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)
225
226
227
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
228

229
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
230

lintangsutawika's avatar
lintangsutawika committed
231
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
232
233
234
        self.fewshot_rnd: Optional[
            random.Random
        ] = None  # purposely induce errors in case of improper usage
235

236
237
238
239
240
241
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
        """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.
        """
266
267
268
269
270
271
272
        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,
        )
273

274
    @property
275
    def config(self) -> TaskConfig:
276
277
278
        """Returns the TaskConfig associated with this class."""
        return self._config

279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
    @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

294
    def training_docs(self) -> Iterable:
295
296
297
298
299
300
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

301
    def validation_docs(self) -> Iterable:
302
303
304
305
306
307
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

308
    def test_docs(self) -> Iterable:
309
310
311
312
313
314
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

315
    def fewshot_docs(self) -> Iterable:
316
317
318
319
320
321
322
323
324
        """
        :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
325
            eval_logger.warning(
326
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
327
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
328
            )
329
330
            return self.test_docs()

331
    def _process_doc(self, doc: dict) -> dict:
332
333
334
335
336
337
338
339
340
        """
        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
341

342
    @property
343
    def instances(self) -> List[Instance]:
344
345
346
347
348
349
350
351
352
353
354
        """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)

355
356
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
357
358
359
360
361
362
363
364
365
366
367
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
    def doc_to_text(self, doc):
        pass

    @abc.abstractmethod
    def doc_to_target(self, doc):
        pass

368
369
    def build_all_requests(
        self,
370
        *,
371
372
373
374
375
        limit=None,
        rank=None,
        world_size=None,
        cache_requests=False,
        rewrite_requests_cache=False,
Konrad's avatar
Konrad committed
376
        system_instruction=None,
Konrad's avatar
Konrad committed
377
        apply_chat_template=False,
Konrad's avatar
Konrad committed
378
        fewshot_as_multiturn=False,
Konrad's avatar
Konrad committed
379
        lm=None,
380
    ) -> None:
381
        """Build a set of Instances for a task, and store them in task.instances"""
382
383
384
385

        # used with caching
        og_limit = limit

386
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401

        cached_instances = load_from_cache(file_name=cache_key)

        if cache_requests and cached_instances and not rewrite_requests_cache:
            cached_instances = cached_instances[:limit]

            flattened_instances = [
                instance
                for instance_group in cached_instances
                for instance in instance_group
            ]

            self._instances = flattened_instances
            return

Baber Abbasi's avatar
Baber Abbasi committed
402
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
403

404
        instances = []
405
406
407
408
409
410
411
412
413
414

        # process all documents when caching is specified for simplicity
        if (
            cache_requests
            and (not cached_instances or rewrite_requests_cache)
            and limit is not None
        ):
            limit = None

        doc_id_docs = list(
415
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
416
417
418
419
420
421
422
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
423
        ):
424
            # sample fewshot context #TODO: need to offset doc_id by rank now!
425
            fewshot_ctx = self.fewshot_context(
426
                doc,
427
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
Konrad's avatar
Konrad committed
428
                system_instruction,
Konrad's avatar
Konrad committed
429
                apply_chat_template,
Konrad's avatar
Konrad committed
430
                fewshot_as_multiturn,
Konrad's avatar
Konrad committed
431
                lm,
432
            )
433

434
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
435
436
437
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
438
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
439
            )
440
441
442
443

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

444
445
446
447
448
449
450
451
452
453
454
455
456
            instances.append(inst)

        # now flatten, this is to allow slicing to work with pickles

        sliced_instances = instances[:og_limit]

        flattened_instances = [
            instance
            for instance_group in sliced_instances
            for instance in instance_group
        ]

        self._instances = flattened_instances
457

458
459
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
460

461
462
463
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
    @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
480
            The number of times each instance in a dataset is inferred on. Defaults to 1,
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
            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

516
517
518
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
519
520
521
522
523
524
525
526
527
528
    @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))

529
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
530
    def fewshot_context(
531
532
533
        self,
        doc,
        num_fewshot,
534
        rnd=None,
535
        description=None,
lintangsutawika's avatar
lintangsutawika committed
536
    ):
537
538
539
540
541
542
543
        """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
544
545
546
547
548
        :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.
549
550
551
        :returns: str
            The fewshot context.
        """
552
        if rnd is None:
553
554
555
556
557
558
            if self.fewshot_rnd is not None:
                rnd = self.fewshot_rnd
            else:
                raise ValueError(
                    "A `random.Random` generator argument must be provided to `rnd`"
                )
lintangsutawika's avatar
lintangsutawika committed
559

560
        description = description if description else ""
561
562

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
563
            labeled_examples = ""
564
        else:
lintangsutawika's avatar
lintangsutawika committed
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
            # 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
589
            )
590
591

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
592
        return description + labeled_examples + example
593

594
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
595
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
596
597
        if hasattr(self, "_filters"):
            for f in self._filters:
598
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
599
600
601
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
602

baberabb's avatar
baberabb committed
603
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
604
        """Returns the config as a dictionary."""
605
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
606
        # (num_fewshot)
607
        return self.config.to_dict()
608

Baber Abbasi's avatar
Baber Abbasi committed
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
    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)

649
650
651
652
653
    def set_fewshot_seed(self, seed: Optional[int] = None) -> None:
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

654
655
656
657
658
659
660
    @property
    def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
661
662
663
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
664
665
666
667
668
669
670
671
672
673
674
675
676

    def doc_iterator(
        self, *, rank: int = 0, limit: Union[int, None] = None, world_size: int = 1
    ) -> Iterator[Tuple[int, Any]]:
        limit = int(limit) if limit else None
        doc_iterator = utils.create_iterator(
            enumerate(self.eval_docs),
            rank=int(rank),
            limit=limit,
            world_size=int(world_size),
        )
        return doc_iterator

677
678

class ConfigurableTask(Task):
679
    VERSION = "Yaml"
680
    OUTPUT_TYPE = None
681
    CONFIG = None
682
683

    def __init__(
684
685
686
687
688
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
689
    ) -> None:  # TODO no super() call here
690
        # Get pre-configured attributes
691
        self._config = self.CONFIG
692

693
        # Use new configurations if there was no preconfiguration
694
        if self.config is None:
695
            self._config = TaskConfig(**config)
696
697
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
698
            if config is not None:
699
                self._config.__dict__.update(config)
700

701
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
702
703
704
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
705

706
707
708
709
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

710
        if self.config.output_type is not None:
711
712
713
714
            if self.config.output_type not in ALL_OUTPUT_TYPES:
                raise ValueError(
                    f"Got invalid output_type '{self.config.output_type}', must be in '{','.join(ALL_OUTPUT_TYPES)}'"
                )
715
            self.OUTPUT_TYPE = self.config.output_type
716

717
718
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
719

720
721
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
722

723
724
725
726
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
727

728
        if self.config.metric_list is None:
729
            # TODO: handle this in TaskConfig.__post_init__ ?
730
731
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

732
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
733
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
734
                self._metric_fn_kwargs[metric_name] = {}
735
736
737
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
738
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
739
        else:
740
            for metric_config in self.config.metric_list:
741
742
743
744
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
745
746
747
748
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
749
750
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
751
                }
Chris's avatar
Chris committed
752
753
754
755
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
756

757
                if self.config.process_results is not None:
758
759
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
760
761
762
763
764
765
                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
766
767
768
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
769
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
770

771
                if "aggregation" in metric_config:
772
                    agg_name = metric_config["aggregation"]
773
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
774
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
775
                    elif callable(agg_name):  # noqa: E721
776
777
778
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
779
                else:
780
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
781
                    metric_agg = get_metric_aggregation(metric_name)
782
                    eval_logger.warning(
783
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
784
785
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
786
                    )
787
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
788

789
790
791
792
793
794
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
795
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
796
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
797
                        f"higher_is_better={is_higher_better(metric_name)}"
798
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
799
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
800

801
        self.download(self.config.dataset_kwargs)
802
803
804
        self._training_docs = None
        self._fewshot_docs = None

805
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
806
            self._filters = []
807
            for filter_config in self.config.filter_list:
808
809
810
811
812
813
814
815
816
                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
817
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
818
        else:
819
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
820

821
822
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
823
            self.prompt = get_prompt(
824
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
825
            )
826
827
828
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
829
        if self.fewshot_docs() is not None:
830
831
832
833
            self.fewshot_rnd = (
                random.Random()
            )  # setting with no seed, to be overridden at a later time
            config_sampler: Union[str, Callable] = (
haileyschoelkopf's avatar
haileyschoelkopf committed
834
835
836
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
            )
            if isinstance(config_sampler, str):
                self.sampler = samplers.get_sampler(config_sampler)(
                    list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
                )
            elif callable(config_sampler) and issubclass(
                config_sampler, samplers.ContextSampler
            ):
                self.sampler = config_sampler(
                    docs=list(self.fewshot_docs()), task=self, rnd=self.fewshot_rnd
                )
            else:
                raise TypeError(
                    f"fewshot_config.sampler should be a string or callable of ContextSampler type, "
                    f"not {type(config_sampler)}"
                )
853

854
        self.task_docs = self.eval_docs
855

856
        # Test One Doc
857
        self.features = list(self.task_docs.features.keys())
858
859
        self.multiple_input = 0
        self.multiple_target = 0
860
        test_doc = self.task_docs[0]
861
        test_text = self.doc_to_text(test_doc)
862
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
863

864
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
865
            test_choice = self.doc_to_choice(test_doc)
866
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
867
                eval_logger.error("doc_to_choice must return list")
868
869
            else:
                num_choice = len(test_choice)
870

871
            if isinstance(test_text, int):
872
                self.multiple_input = num_choice
873
874
        else:
            test_choice = None
875

876
        if isinstance(test_target, list):
877
            self.multiple_target = len(test_target)
878
        else:
879
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
880
                test_target = test_choice[test_target]
881
            else:
lintangsutawika's avatar
lintangsutawika committed
882
                test_target = str(test_target)
883

884
885
886
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
887
            check_choices = [test_target]
888
889
890
891
        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 = (
892
893
                    True
                    if self.config.target_delimiter.rstrip()
894
                    != self.config.target_delimiter
895
                    else False
896
                )
897

898
                if delimiter_has_whitespace and choice_has_whitespace:
899
900
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
901
902
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
903
                    eval_logger.debug(
904
                        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'
905
906
                    )

907
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
908
909
910
911
912
913
        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
914
    def has_training_docs(self) -> bool:
915
        if self.config.training_split is not None:
916
917
918
919
            return True
        else:
            return False

baberabb's avatar
baberabb committed
920
    def has_validation_docs(self) -> bool:
921
        if self.config.validation_split is not None:
922
923
924
925
            return True
        else:
            return False

baberabb's avatar
baberabb committed
926
    def has_test_docs(self) -> bool:
927
        if self.config.test_split is not None:
928
929
930
931
            return True
        else:
            return False

baberabb's avatar
baberabb committed
932
    def training_docs(self) -> datasets.Dataset:
933
        if self.has_training_docs():
934
935
936
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
937
                )
938
            return self.dataset[self.config.training_split]
939

baberabb's avatar
baberabb committed
940
    def validation_docs(self) -> datasets.Dataset:
941
        if self.has_validation_docs():
942
943
944
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
945
                )
946
            return self.dataset[self.config.validation_split]
947

baberabb's avatar
baberabb committed
948
    def test_docs(self) -> datasets.Dataset:
949
        if self.has_test_docs():
950
951
952
            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]
953

954
    def fewshot_docs(self):
955
        if self.config.fewshot_split is not None:
956
957
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
958
            return self.dataset[self.config.fewshot_split]
959
        else:
960
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
961
                eval_logger.warning(
962
                    f"Task '{self.config.task}': "
963
964
965
966
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
967

Konrad's avatar
Konrad committed
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
    @staticmethod
    def append_target_question(
        labeled_examples: list[dict[str, str]],
        question: str,
        fewshot_as_multiturn: bool = False,
    ) -> None:
        """Adds a target question to the labeled examples list.
        If fewshot_as_multiturn is True, or labeled_examples is empty, or the last entry is a system turn, appends the question as a new user entry.
        Otherwise, it is appended to the last user entry, ensuring that the conversation alternates between the user and the assistant.
        """
        if not fewshot_as_multiturn:
            # if no messages or last message is system, append as new user entry
            if len(labeled_examples) == 0 or labeled_examples[-1]["role"] == "system":
                labeled_examples.append({"role": "user", "content": question})
            # if last message is user, append to it to avoid two user messages in a row
            else:
                labeled_examples[-1]["content"] += question
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
            labeled_examples.append({"role": "user", "content": question})

lintangsutawika's avatar
lintangsutawika committed
989
    @utils.positional_deprecated
Konrad's avatar
Konrad committed
990
991
992
993
    def fewshot_context(
        self,
        doc: str,
        num_fewshot: int,
Konrad's avatar
Konrad committed
994
        system_instruction: Optional[str] = None,
Konrad's avatar
Konrad committed
995
        apply_chat_template: bool = False,
Konrad's avatar
Konrad committed
996
        fewshot_as_multiturn: bool = False,
Konrad's avatar
Konrad committed
997
        lm=None,
Konrad's avatar
Konrad committed
998
    ) -> str:
lintangsutawika's avatar
lintangsutawika committed
999
1000
1001
1002
1003
1004
1005
        """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.
Konrad's avatar
Konrad committed
1006
1007
        :param  system_instruction: str
            System instruction to be applied to the prompt.
Konrad's avatar
Konrad committed
1008
1009
        :param apply_chat_template: bool
            Whether to apply the chat template to the fewshot context.
Konrad's avatar
Konrad committed
1010
1011
        :param fewshot_as_multiturn: bool
            Whether to provide the fewshot examples as a multiturn conversation or a single user turn.
Konrad's avatar
Konrad committed
1012
1013
        :param lm:
            Language model with definition of the tokenizer/function to use for applying the chat template.
lintangsutawika's avatar
lintangsutawika committed
1014
1015
1016
        :returns: str
            The fewshot context.
        """
Konrad's avatar
Konrad committed
1017
1018
1019
1020
1021
1022
1023

        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1024
1025
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1026

Konrad's avatar
Konrad committed
1027
        # create system prompt based on the provided system instruction and description
Konrad's avatar
Konrad committed
1028
        if system_instruction is not None and description:
Konrad's avatar
Konrad committed
1029
1030
1031
            system_prompt = (
                f"{system_instruction}{self.sampler.fewshot_delimiter}{description}"
            )
Konrad's avatar
Konrad committed
1032
        elif system_instruction is not None:
Konrad's avatar
Konrad committed
1033
1034
1035
1036
1037
1038
1039
1040
            system_prompt = system_instruction
        elif description:
            system_prompt = description
        else:
            system_prompt = ""

        # add system prompt if specified
        if system_prompt:
Konrad's avatar
Konrad committed
1041
            if apply_chat_template:
Konrad's avatar
Konrad committed
1042
                labeled_examples.append({"role": "system", "content": system_prompt})
Konrad's avatar
Konrad committed
1043
            else:
Konrad's avatar
Konrad committed
1044
1045
1046
1047
                labeled_examples = system_prompt

        # if few-shot - append examples after the system prompt
        if num_fewshot > 0:
Konrad's avatar
Konrad committed
1048
            if apply_chat_template:
1049
1050
1051
1052
                labeled_examples.extend(
                    self.sampler.get_chat_context(
                        doc, num_fewshot, fewshot_as_multiturn
                    )
Konrad's avatar
Konrad committed
1053
1054
                )
            else:
Konrad's avatar
Konrad committed
1055
                labeled_examples += self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1056
1057

        example = self.doc_to_text(doc)
Konrad's avatar
Konrad committed
1058
        if apply_chat_template:
Konrad's avatar
Konrad committed
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
            if self.multiple_input:
                return lm.apply_chat_template(labeled_examples)
            if isinstance(example, str):
                self.append_target_question(
                    labeled_examples, example, fewshot_as_multiturn
                )
            # for loglikelihood create a list of questions with appended choices
            elif isinstance(example, list):
                labeled_examples_list = []
                # copy chat history for each example and append the answer
                for ex in example:
                    chat = deepcopy(labeled_examples)
                    self.append_target_question(chat, ex, fewshot_as_multiturn)
                    labeled_examples_list.append(lm.apply_chat_template(chat))
                return labeled_examples_list
            # if example is an integer, append the choice or convert to string
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
                    self.append_target_question(
                        labeled_examples, choices[example], fewshot_as_multiturn
                    )
                else:
                    self.append_target_question(
                        labeled_examples, str(example), fewshot_as_multiturn
                    )
                # return lm.apply_chat_template(labeled_examples)
Konrad's avatar
Konrad committed
1086
            return lm.apply_chat_template(labeled_examples)
1087
        else:
Konrad's avatar
Konrad committed
1088
1089
            if self.multiple_input:
                return labeled_examples
Konrad's avatar
Konrad committed
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
            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
1100

1101
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1102
        """Iterates over FilterEnsembles and applies them to instances"""
1103
1104
        if hasattr(self, "_filters"):
            for f in self._filters:
1105
                f.apply(self._instances)
1106
1107
1108
1109
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1110
    def should_decontaminate(self):
1111
        return self.config.should_decontaminate
1112
1113

    def doc_to_decontamination_query(self, doc):
1114
        if self.config.should_decontaminate:
1115
1116
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1117
            else:
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
                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
                        )
                    )
1129

1130
    def _process_doc(self, doc: dict) -> dict:
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
        """
        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):
1142
1143
        if self.prompt is not None:
            doc_to_text = self.prompt
1144
        else:
1145
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1146

1147
        if isinstance(doc_to_text, int):
1148
            return doc_to_text
1149
        elif isinstance(doc_to_text, str):
1150
            if doc_to_text in self.features:
1151
                # if self.config.doc_to_choice is not None:
1152
1153
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1154
1155
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1156
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1157
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1158
1159
1160
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1161
        elif callable(doc_to_text):
1162
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1163
        # Used when applying a Promptsource template
1164
        elif hasattr(doc_to_text, "apply"):
1165
1166
1167
1168
1169
            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")
1170
                return self.config.fewshot_delimiter
1171
        else:
1172
            print(type(doc_to_text))
1173
            raise TypeError
1174

1175
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1176
1177
        if self.prompt is not None:
            doc_to_target = self.prompt
1178
        else:
1179
            doc_to_target = self.config.doc_to_target
1180

1181
        if isinstance(doc_to_target, int):
1182
            return doc_to_target
1183
        elif isinstance(doc_to_target, str):
1184
            if doc_to_target in self.features:
1185
                # if self.config.doc_to_choice is not None:
1186
1187
1188
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1189
            else:
lintangsutawika's avatar
lintangsutawika committed
1190
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1191
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1192
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1193
1194
1195
1196
1197
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1198
1199
1200
1201
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1202
1203
                else:
                    return target_string
1204
        elif isinstance(doc_to_target, list):
1205
            return doc_to_target
1206
        elif callable(doc_to_target):
1207
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1208
        # Used when applying a Promptsource template
1209
        elif hasattr(doc_to_target, "apply"):
1210
            applied_prompt = doc_to_target.apply(doc)
1211
1212
1213
1214
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1215
                return self.config.fewshot_delimiter
1216
1217
        else:
            raise TypeError
1218

baberabb's avatar
baberabb committed
1219
    def doc_to_choice(self, doc: Any) -> List[str]:
1220
1221
        if self.prompt is not None:
            doc_to_choice = self.prompt
1222
        elif self.config.doc_to_choice is None:
1223
1224
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1225
            doc_to_choice = self.config.doc_to_choice
1226

1227
        if isinstance(doc_to_choice, str):
1228
1229
1230
1231
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1232
        elif isinstance(doc_to_choice, list):
1233
            return doc_to_choice
1234
        elif isinstance(doc_to_choice, dict):
1235
1236
1237
1238
1239
1240
1241
            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
1242

baberabb's avatar
baberabb committed
1243
1244
1245
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1246
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1247
            arguments = (ctx, self.doc_to_target(doc))
1248
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1249
            arguments = (self.doc_to_target(doc),)
1250
        elif self.OUTPUT_TYPE == "multiple_choice":
1251
            choices = self.doc_to_choice(doc)
1252
            target_delimiter = self.config.target_delimiter
1253
1254
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1255
                cont = self.doc_to_target(doc)
1256
1257
1258
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1259
            else:
1260
                # Otherwise they are placed in the continuation
1261
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1262

1263
            request_list = [
1264
1265
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1266
                    doc=doc,
1267
                    arguments=arg,
1268
                    idx=i,
1269
1270
                    **kwargs,
                )
1271
                for i, arg in enumerate(arguments)
1272
            ]
1273
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1274
            if "acc_mutual_info" in self._metric_fn_list.keys():
1275
1276
1277
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1278
                # here mutual info refers to calculating
1279
1280
1281
1282
1283
1284
                # 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
1285
                            doc=doc,
1286
                            arguments=("", "{}".format(choice)),
1287
1288
1289
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1290
                        for i, choice in enumerate(choices)
1291
1292
1293
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1294

1295
        elif self.OUTPUT_TYPE == "generate_until":
1296
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1297
1298

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1299
1300
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1301
1302

    def process_results(self, doc, results):
1303
1304
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1305

1306
        result_dict = {}
1307
        use_metric = list(self._metric_fn_list.keys())
1308
1309
1310
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1311
1312
1313
1314
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1315
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1316
            (loglikelihood,) = results
1317
1318
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1319
            return {
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
                **(
                    {"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
1335
            }
1336
        elif self.OUTPUT_TYPE == "multiple_choice":
1337
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1338

1339
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1340
            choices = self.doc_to_choice(doc)
1341
1342
            completion_len = np.array([float(len(i)) for i in choices])

1343
1344
            if (
                2 * len(choices) == len(lls)
1345
                and "acc_mutual_info" in self._metric_fn_list.keys()
1346
1347
1348
1349
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1350
1351
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1352
1353
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1354

1355
1356
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1357

1358
1359
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1360
            else:
1361
                gold = self.doc_to_target(doc)
1362
1363

            gold_index_error = False
1364
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1365
1366
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1367
1368
                    gold_index_error = True
            else:
1369
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1370
                    gold = gold if gold < len(choices) else -100
1371
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1372
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1373

Lintang Sutawika's avatar
Lintang Sutawika committed
1374
                if gold == -100:
1375
1376
1377
1378
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1379
                    f"Label index was not in within range of available choices,"
1380
1381
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1382

1383
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1384
1385
                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
1386
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1387
1388
1389
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1390
                # 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
1391
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1392

Lintang Sutawika's avatar
Lintang Sutawika committed
1393
1394
1395
1396
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1397
            result_dict = {
1398
                **({"acc": acc} if "acc" in use_metric else {}),
1399
1400
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1401
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1402
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1403
1404
1405
1406
1407
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1408
1409
            }

1410
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1411
1412
1413
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1414
1415
1416
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1417
        elif self.OUTPUT_TYPE == "generate_until":
1418
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1419
            result = results[0]
1420
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1421
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1422
                # it assumes that doc_to_target returns a number.
1423
1424
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1425
1426
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1427
                gold = list(gold)
Chris's avatar
Chris committed
1428
1429
1430
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1431

lintangsutawika's avatar
lintangsutawika committed
1432
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1433
1434
1435
1436
1437
                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
1438
1439
1440
1441
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1442
1443
1444
1445
1446
1447
1448
1449
                    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
1450
                    else:
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
                        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
1472
                else:
1473
                    try:
1474
                        result_score = self._metric_fn_list[metric](
1475
1476
                            references=[gold],
                            predictions=[result],
1477
                            **self._metric_fn_kwargs[metric],
1478
                        )
1479
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1480
                        result_score = self._metric_fn_list[metric]([gold, result])
1481
1482
1483
1484
                    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
1485
        else:
lintangsutawika's avatar
lintangsutawika committed
1486
1487
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1488
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1489
            )
1490
1491
1492

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1493
    def aggregation(self) -> dict:
1494
1495
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1496
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1497
        return self._higher_is_better
1498

Baber Abbasi's avatar
Baber Abbasi committed
1499
1500
1501
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1502
1503
1504
1505
1506
1507
1508
1509
1510
    def __repr__(self):
        return (
            f"ConfigurableTask(task_name={getattr(self.config, 'task', None)},"
            f"group_name={getattr(self.config, 'group', None)},"
            f"output_type={self.OUTPUT_TYPE},"
            f"num_fewshot={getattr(self.config, 'num_fewshot', None)},"
            f"num_samples={len(self.eval_docs)})"
        )

1511
1512

class MultipleChoiceTask(Task):
1513
    OUTPUT_TYPE = "loglikelihood"
1514

baberabb's avatar
baberabb committed
1515
    def doc_to_target(self, doc: dict) -> str:
1516
1517
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1518
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1519
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1520
1521
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1522
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1523
                doc=doc,
1524
                arguments=(ctx, " {}".format(choice)),
1525
                idx=i,
1526
1527
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1528
1529
            for i, choice in enumerate(doc["choices"])
        ]
1530

1531
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1532
1533
1534
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
        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
1546
    def higher_is_better(self) -> dict:
1547
1548
1549
1550
1551
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1552
    def aggregation(self) -> dict:
1553
1554
1555
1556
1557
1558
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1559
class PerplexityTask(Task):
1560
1561
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1562
    def has_training_docs(self) -> bool:
1563
1564
        return False

baberabb's avatar
baberabb committed
1565
    def fewshot_examples(self, k: int, rnd) -> List:
1566
1567
1568
1569
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1570
1571
        return []

baberabb's avatar
baberabb committed
1572
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1573
1574
1575
1576
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1577
1578
1579

        return ""

baberabb's avatar
baberabb committed
1580
    def higher_is_better(self) -> dict:
1581
1582
1583
1584
1585
1586
1587
1588
1589
        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
1590
    def doc_to_text(self, doc) -> str:
1591
1592
1593
1594
1595
        return ""

    def doc_to_target(self, doc):
        return doc

1596
1597
1598
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1599

lintangsutawika's avatar
lintangsutawika committed
1600
1601
1602
1603
1604
1605
1606
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1607

1608
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1609
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1610
1611
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1612
1613
1614
1615
1616
1617
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1618
    def aggregation(self) -> dict:
1619
1620
1621
1622
1623
1624
1625
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1626
    def count_bytes(cls, doc) -> int:
1627
1628
1629
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1630
    def count_words(cls, doc) -> int:
1631
1632
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))