task.py 65.5 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
8
from dataclasses import asdict, dataclass
9
from inspect import getsource
10
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
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
47
    "multiple_choice_gpt",
48
    "loglikelihood_rolling",
49
    "generate_until",
50
51
]

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

lintangsutawika's avatar
lintangsutawika committed
54

55
56
@dataclass
class TaskConfig(dict):
57
    # task naming/registry
58
59
60
61
    task: Optional[str] = None
    task_alias: Optional[str] = None
    group: Optional[Union[str, list]] = None
    group_alias: Optional[Union[str, list]] = None
62
63
64
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
65
66
67
68
69
70
    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
71
72
73
    fewshot_split: Optional[str] = (
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaling (?)
    )
74
75
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
76
77
78
79
80
81
    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
82
    description: str = ""
83
84
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
85
    fewshot_config: Optional[dict] = None
86
    # runtime configuration options
87
    num_fewshot: Optional[int] = None
88
    # scoring options
89
90
91
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
92
    repeats: int = 1
93
    filter_list: Optional[Union[str, list]] = None
94
    should_decontaminate: bool = False
95
    doc_to_decontamination_query: Optional[str] = None
96
97
98
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
99

Ethan Smith's avatar
Ethan Smith committed
100
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
101
        if self.generation_kwargs is not None:
102
            if self.output_type != "generate_until":
103
                eval_logger.warning(
104
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
105
106
107
108
109
110
111
112
                )

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

            if "until" not in self.generation_kwargs:
113
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
114
        else:
115
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
116
117
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
118
119
120
121
122
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
123
124
                    "do_sample": False,
                }
125

126
127
128
    def __getitem__(self, item):
        return getattr(self, item)

129
130
131
    def __setitem__(self, item, value):
        return setattr(self, item, value)

132
    def to_dict(self, keep_callable: bool = False) -> dict:
133
134
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
135
        Used for dumping results alongside full task configuration
136

haileyschoelkopf's avatar
haileyschoelkopf committed
137
138
139
140
141
142
143
144
145
146
        :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)
147
148
149
150
151
152
153
154
155
156
            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
157
        return cfg_dict
158

159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
    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)

175
176
177
178
179
180
181
182
183
184
185

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)
    """

186
    VERSION: Optional[Union[int, str]] = None
187

188
189
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
190
    DATASET_PATH: Optional[str] = None
191
192

    # The name of a subset within `DATASET_PATH`.
193
    DATASET_NAME: Optional[str] = None
194

195
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
196

197
198
    def __init__(
        self,
199
200
201
202
        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
203
    ) -> None:
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
        """
        :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)
226
227
228
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
229

230
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
231

lintangsutawika's avatar
lintangsutawika committed
232
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
233
234
235
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
236

237
238
239
240
241
242
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
        """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.
        """
267
268
269
270
271
272
273
        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,
        )
274

275
    @property
276
    def config(self) -> TaskConfig:
277
278
279
        """Returns the TaskConfig associated with this class."""
        return self._config

280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
    @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

295
    def training_docs(self) -> Iterable:
296
297
298
299
300
301
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

302
    def validation_docs(self) -> Iterable:
303
304
305
306
307
308
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

309
    def test_docs(self) -> Iterable:
310
311
312
313
314
315
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

316
    def fewshot_docs(self) -> Iterable:
317
318
319
320
321
322
323
324
325
        """
        :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
326
            eval_logger.warning(
327
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
328
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
329
            )
330
331
            return self.test_docs()

332
    def _process_doc(self, doc: dict) -> dict:
333
334
335
336
337
338
339
340
341
        """
        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
342

343
    @property
344
    def instances(self) -> List[Instance]:
345
346
347
348
349
350
351
352
353
354
355
        """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)

356
357
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
358
359
360
361
362
363
364
365
366
367
368
            "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

369
370
    def build_all_requests(
        self,
371
        *,
372
373
374
375
376
377
378
379
380
381
        limit: Union[int, None] = None,
        rank: int = 0,
        world_size: int = 1,
        cache_requests: bool = False,
        rewrite_requests_cache: bool = False,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
        chat_template: Optional[Callable] = None,
        tokenizer_name: str = "",
382
    ) -> None:
383
        """Build a set of Instances for a task, and store them in task.instances"""
384
385
386
387

        # used with caching
        og_limit = limit

388
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
389
390
391
392
393
394
395
        cache_key += "-chat_template" if apply_chat_template else ""
        cache_key += "-fewshot_as_multiturn" if fewshot_as_multiturn else ""
        cache_key += (
            f"-system_prompt_hash{utils.hash_string(system_instruction)}"
            if system_instruction is not None
            else ""
        )
396
        cache_key += f"-tokenizer{tokenizer_name}"
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411

        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
412
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
413

414
        instances = []
415
416
417
418
419
420
421
422
423
424

        # 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(
425
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
426
427
428
429
430
431
432
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
433
        ):
434
            # sample fewshot context #TODO: need to offset doc_id by rank now!
435
            fewshot_ctx = self.fewshot_context(
436
                doc,
437
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
438
439
440
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
441
                chat_template,
442
            )
443

444
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
445
446
447
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
448
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
449
            )
450
451
452
453

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

454
455
456
457
458
459
460
461
462
463
464
465
466
            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
467

468
469
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
470

471
472
473
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
    @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
490
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
516
517
518
519
520
521
522
523
524
525
            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

526
527
528
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
529
530
531
532
533
534
535
536
537
538
    @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))

539
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
540
    def fewshot_context(
541
542
543
        self,
        doc,
        num_fewshot,
544
        rnd=None,
545
        description=None,
lintangsutawika's avatar
lintangsutawika committed
546
    ):
547
548
549
550
551
552
553
        """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
554
555
556
557
558
        :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.
559
560
561
        :returns: str
            The fewshot context.
        """
562
        if rnd is None:
563
564
565
566
567
568
            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
569

570
        description = description if description else ""
571
572

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
573
            labeled_examples = ""
574
        else:
lintangsutawika's avatar
lintangsutawika committed
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
            # 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
599
            )
600
601

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
602
        return description + labeled_examples + example
603

604
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
605
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
606
607
        if hasattr(self, "_filters"):
            for f in self._filters:
608
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
609
610
611
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
612

baberabb's avatar
baberabb committed
613
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
614
        """Returns the config as a dictionary."""
615
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
616
        # (num_fewshot)
617
        return self.config.to_dict()
618

Baber Abbasi's avatar
Baber Abbasi committed
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
649
650
651
652
653
654
655
656
657
658
    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)

659
660
661
662
663
    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

664
665
666
667
668
669
670
    @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:
671
672
673
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
674
675
676
677
678
679
680
681
682
683
684
685
686

    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

687
688

class ConfigurableTask(Task):
689
    VERSION = "Yaml"
690
    OUTPUT_TYPE = None
691
    CONFIG = None
692
693

    def __init__(
694
695
696
697
698
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
699
    ) -> None:  # TODO no super() call here
700
        # Get pre-configured attributes
701
        self._config = self.CONFIG
702

703
        # Use new configurations if there was no preconfiguration
704
        if self.config is None:
705
            self._config = TaskConfig(**config)
706
707
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
708
            if config is not None:
709
                self._config.__dict__.update(config)
710

711
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
712
713
714
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
715

716
717
718
719
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

720
        if self.config.output_type is not None:
721
722
723
724
            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)}'"
                )
725
            self.OUTPUT_TYPE = self.config.output_type
726

727
728
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
729

730
731
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
732

733
734
735
736
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
737

738
        if self.config.metric_list is None:
739
            # TODO: handle this in TaskConfig.__post_init__ ?
740
741
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

742
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
743
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
744
                self._metric_fn_kwargs[metric_name] = {}
745
746
747
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
748
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
749
        else:
750
            for metric_config in self.config.metric_list:
751
752
753
754
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
755
756
757
758
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
759
760
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
761
                }
Chris's avatar
Chris committed
762
763
764
765
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
766

767
                if self.config.process_results is not None:
768
769
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
770
771
772
773
774
775
                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
776
777
778
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
779
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
780

781
                if "aggregation" in metric_config:
782
                    agg_name = metric_config["aggregation"]
783
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
784
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
785
                    elif callable(agg_name):  # noqa: E721
786
787
788
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
789
                else:
790
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
791
                    metric_agg = get_metric_aggregation(metric_name)
792
                    eval_logger.warning(
793
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
794
795
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
796
                    )
797
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
798

799
800
801
802
803
804
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
805
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
806
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
807
                        f"higher_is_better={is_higher_better(metric_name)}"
808
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
809
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
810

811
        self.download(self.config.dataset_kwargs)
812
813
814
        self._training_docs = None
        self._fewshot_docs = None

815
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
816
            self._filters = []
817
            for filter_config in self.config.filter_list:
818
819
820
821
822
823
824
825
826
                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
827
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
828
        else:
829
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
830

831
832
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
833
            self.prompt = get_prompt(
834
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
835
            )
836
837
838
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
839
        if self.fewshot_docs() is not None:
840
841
842
843
            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
844
845
846
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
            )
            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)}"
                )
863

864
        self.task_docs = self.eval_docs
865

866
        # Test One Doc
867
        self.features = list(self.task_docs.features.keys())
868
869
        self.multiple_input = 0
        self.multiple_target = 0
870
        test_doc = self.task_docs[0]
871
        test_text = self.doc_to_text(test_doc)
872
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
873

874
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
875
            test_choice = self.doc_to_choice(test_doc)
876
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
877
                eval_logger.error("doc_to_choice must return list")
878
879
            else:
                num_choice = len(test_choice)
880

881
            if isinstance(test_text, int):
882
                self.multiple_input = num_choice
883
884
        else:
            test_choice = None
885

886
        if isinstance(test_target, list):
887
            self.multiple_target = len(test_target)
888
        else:
889
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
890
                test_target = test_choice[test_target]
891
            else:
lintangsutawika's avatar
lintangsutawika committed
892
                test_target = str(test_target)
893

894
895
896
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
897
            check_choices = [test_target]
898
899
900
901
        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 = (
902
903
                    True
                    if self.config.target_delimiter.rstrip()
904
                    != self.config.target_delimiter
905
                    else False
906
                )
907

908
                if delimiter_has_whitespace and choice_has_whitespace:
909
910
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
911
912
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
913
                    eval_logger.debug(
914
                        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'
915
916
                    )

917
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
918
919
920
921
922
923
        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
924
    def has_training_docs(self) -> bool:
925
        if self.config.training_split is not None:
926
927
928
929
            return True
        else:
            return False

baberabb's avatar
baberabb committed
930
    def has_validation_docs(self) -> bool:
931
        if self.config.validation_split is not None:
932
933
934
935
            return True
        else:
            return False

baberabb's avatar
baberabb committed
936
    def has_test_docs(self) -> bool:
937
        if self.config.test_split is not None:
938
939
940
941
            return True
        else:
            return False

baberabb's avatar
baberabb committed
942
    def training_docs(self) -> datasets.Dataset:
943
        if self.has_training_docs():
944
945
946
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
947
                )
948
            return self.dataset[self.config.training_split]
949

baberabb's avatar
baberabb committed
950
    def validation_docs(self) -> datasets.Dataset:
951
        if self.has_validation_docs():
952
953
954
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
955
                )
956
            return self.dataset[self.config.validation_split]
957

baberabb's avatar
baberabb committed
958
    def test_docs(self) -> datasets.Dataset:
959
        if self.has_test_docs():
960
961
962
            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]
963

964
    def fewshot_docs(self):
965
        if self.config.fewshot_split is not None:
966
967
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
968
            return self.dataset[self.config.fewshot_split]
969
970
971
972
973
974
975
976
977
978
979
980
        elif (
            self.config.fewshot_config is not None
            and self.config.fewshot_config.get("samples", None) is not None
        ):
            if isinstance(self.config.fewshot_config["samples"], list):
                return self.config.fewshot_config["samples"]
            elif callable(self.config.fewshot_config["samples"]):
                return self.config.fewshot_config["samples"]()
            else:
                raise Exception(
                    "`fewshot_config['samples']` was incorrectly defined in the configuration. It should be either a list of samples as a dict, or function returning this list."
                )
981
        else:
982
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
983
                eval_logger.warning(
984
                    f"Task '{self.config.task}': "
985
986
987
988
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
989

KonradSzafer's avatar
KonradSzafer committed
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
    @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
1011
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1012
1013
1014
1015
1016
1017
1018
    def fewshot_context(
        self,
        doc: str,
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1019
        chat_template: Optional[Callable] = None,
KonradSzafer's avatar
KonradSzafer committed
1020
    ) -> str:
lintangsutawika's avatar
lintangsutawika committed
1021
1022
1023
1024
1025
1026
1027
        """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.
KonradSzafer's avatar
KonradSzafer committed
1028
1029
1030
1031
1032
1033
        :param  system_instruction: str
            System instruction to be applied to the prompt.
        :param apply_chat_template: bool
            Whether to apply the chat template to the fewshot context.
        :param fewshot_as_multiturn: bool
            Whether to provide the fewshot examples as a multiturn conversation or a single user turn.
1034
1035
        :param chat_template: Callable
            Chat template to be applied to the fewshot context.
lintangsutawika's avatar
lintangsutawika committed
1036
1037
1038
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1039
1040
1041
1042
1043
1044
1045

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

        # get task description
1046
1047
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1048

KonradSzafer's avatar
KonradSzafer committed
1049
1050
1051
1052
1053
1054
1055
1056
1057
        # create system prompt based on the provided system instruction and description
        if system_instruction is not None and description:
            system_prompt = (
                f"{system_instruction}{self.sampler.fewshot_delimiter}{description}"
            )
        elif system_instruction is not None:
            system_prompt = system_instruction
        elif description:
            system_prompt = description
lintangsutawika's avatar
lintangsutawika committed
1058
        else:
KonradSzafer's avatar
KonradSzafer committed
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
            system_prompt = ""

        # add system prompt if specified
        if system_prompt:
            if apply_chat_template:
                labeled_examples.append({"role": "system", "content": system_prompt})
            else:
                labeled_examples = system_prompt

        # if few-shot - append examples after the system prompt
        if num_fewshot > 0:
            if apply_chat_template:
                labeled_examples.extend(
                    self.sampler.get_chat_context(
                        doc, num_fewshot, fewshot_as_multiturn
                    )
                )
            else:
                labeled_examples += self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1078
1079

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1080
1081
        if apply_chat_template:
            if self.multiple_input:
1082
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
            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)
1094
                    labeled_examples_list.append(chat_template(chat))
KonradSzafer's avatar
KonradSzafer committed
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
                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)
1108
            return chat_template(labeled_examples)
1109
        else:
KonradSzafer's avatar
KonradSzafer committed
1110
1111
            if self.multiple_input:
                return labeled_examples
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
            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
1122

1123
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1124
        """Iterates over FilterEnsembles and applies them to instances"""
1125
1126
        if hasattr(self, "_filters"):
            for f in self._filters:
1127
                f.apply(self._instances)
1128
1129
1130
1131
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1132
    def should_decontaminate(self):
1133
        return self.config.should_decontaminate
1134
1135

    def doc_to_decontamination_query(self, doc):
1136
        if self.config.should_decontaminate:
1137
1138
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1139
            else:
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
                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
                        )
                    )
1151

1152
    def _process_doc(self, doc: dict) -> dict:
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
        """
        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):
1164
1165
        if self.prompt is not None:
            doc_to_text = self.prompt
1166
        else:
1167
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1168

1169
        if isinstance(doc_to_text, int):
1170
            return doc_to_text
1171
        elif isinstance(doc_to_text, str):
1172
            if doc_to_text in self.features:
1173
                # if self.config.doc_to_choice is not None:
1174
1175
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1176
1177
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1178
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1179
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1180
1181
1182
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1183
        elif callable(doc_to_text):
1184
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1185
        # Used when applying a Promptsource template
1186
        elif hasattr(doc_to_text, "apply"):
1187
1188
1189
1190
1191
            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")
1192
                return self.config.fewshot_delimiter
1193
1194
        else:
            raise TypeError
1195

1196
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1197
1198
        if self.prompt is not None:
            doc_to_target = self.prompt
1199
        else:
1200
            doc_to_target = self.config.doc_to_target
1201

1202
        if isinstance(doc_to_target, int):
1203
            return doc_to_target
1204
        elif isinstance(doc_to_target, str):
1205
            if doc_to_target in self.features:
1206
                # if self.config.doc_to_choice is not None:
1207
1208
1209
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1210
            else:
lintangsutawika's avatar
lintangsutawika committed
1211
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1212
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1213
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1214
1215
1216
1217
1218
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1219
1220
1221
1222
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1223
1224
                else:
                    return target_string
1225
        elif isinstance(doc_to_target, list):
1226
            return doc_to_target
1227
        elif callable(doc_to_target):
1228
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1229
        # Used when applying a Promptsource template
1230
        elif hasattr(doc_to_target, "apply"):
1231
            applied_prompt = doc_to_target.apply(doc)
1232
1233
1234
1235
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1236
                return self.config.fewshot_delimiter
1237
1238
        else:
            raise TypeError
1239

baberabb's avatar
baberabb committed
1240
    def doc_to_choice(self, doc: Any) -> List[str]:
1241
1242
        if self.prompt is not None:
            doc_to_choice = self.prompt
1243
        elif self.config.doc_to_choice is None:
1244
1245
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1246
            doc_to_choice = self.config.doc_to_choice
1247

1248
        if isinstance(doc_to_choice, str):
1249
1250
1251
1252
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1253
        elif isinstance(doc_to_choice, list):
1254
            return doc_to_choice
1255
        elif isinstance(doc_to_choice, dict):
1256
1257
1258
1259
1260
1261
1262
            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
1263

baberabb's avatar
baberabb committed
1264
1265
1266
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1267
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1268
            arguments = (ctx, self.doc_to_target(doc))
1269
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1270
            arguments = (self.doc_to_target(doc),)
1271
        elif "multiple_choice" in self.OUTPUT_TYPE:
1272
            choices = self.doc_to_choice(doc)
1273
            target_delimiter = self.config.target_delimiter
1274
1275
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1276
                cont = self.doc_to_target(doc)
1277
1278
1279
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1280
            else:
1281
                # Otherwise they are placed in the continuation
1282
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
            if self.OUTPUT_TYPE == "multiple_choice_gpt":
                request_list = [
                    Instance(
                        request_type="multiple_choice_gpt",
                        doc=doc,
                        arguments=arg,
                        idx=i,
                        **kwargs,
                    )
                    for i, arg in enumerate(arguments)
                ]
            else:
                request_list = [
                    Instance(
                        request_type="loglikelihood",
                        doc=doc,
                        arguments=arg,
                        idx=i,
                        **kwargs,
                    )
                    for i, arg in enumerate(arguments)
                ]
1305
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1306
            if "acc_mutual_info" in self._metric_fn_list.keys():
1307
1308
1309
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1310
                # here mutual info refers to calculating
1311
1312
1313
1314
1315
1316
                # 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
1317
                            doc=doc,
1318
                            arguments=("", "{}".format(choice)),
1319
1320
1321
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1322
                        for i, choice in enumerate(choices)
1323
1324
1325
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1326

1327
        elif self.OUTPUT_TYPE == "generate_until":
1328
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1329
1330

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1331
1332
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1333
1334

    def process_results(self, doc, results):
1335
1336
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1337

1338
        result_dict = {}
1339
        use_metric = list(self._metric_fn_list.keys())
1340
1341
1342
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1343
1344
1345
1346
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1347
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1348
            (loglikelihood,) = results
1349
1350
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1351
            return {
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
                **(
                    {"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
1367
            }
1368
        elif self.OUTPUT_TYPE == "multiple_choice":
1369
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1370

1371
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1372
            choices = self.doc_to_choice(doc)
1373
1374
            completion_len = np.array([float(len(i)) for i in choices])

1375
1376
            if (
                2 * len(choices) == len(lls)
1377
                and "acc_mutual_info" in self._metric_fn_list.keys()
1378
1379
1380
1381
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1382
1383
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1384
1385
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1386

1387
1388
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1389

1390
1391
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1392
            else:
1393
                gold = self.doc_to_target(doc)
1394
1395

            gold_index_error = False
1396
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1397
1398
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1399
1400
                    gold_index_error = True
            else:
1401
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1402
                    gold = gold if gold < len(choices) else -100
1403
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1404
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1405

Lintang Sutawika's avatar
Lintang Sutawika committed
1406
                if gold == -100:
1407
1408
1409
1410
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1411
                    f"Label index was not in within range of available choices,"
1412
1413
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1414

1415
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1416
1417
                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
1418
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1419
1420
1421
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1422
                # 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
1423
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1424

Lintang Sutawika's avatar
Lintang Sutawika committed
1425
1426
1427
1428
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1429
            result_dict = {
1430
                **({"acc": acc} if "acc" in use_metric else {}),
1431
1432
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1433
                **({"squad": (gold, pred)} if "squad" in use_metric else {}),
1434
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1435
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1436
1437
1438
1439
1440
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1441
1442
            }

1443
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1444
1445
1446
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1447
1448
1449
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
        elif self.OUTPUT_TYPE == "multiple_choice_gpt":
            gold = self.doc_to_target(doc)
            result = results[0]
            choices = self.doc_to_choice(doc)
            try:
                gold = choices[gold]
                gold = type(result)(gold)
            except TypeError:
                gold = gold

            for metric in self._metric_fn_list.keys():
                try:
                    result_score = self._metric_fn_list[metric](
                        references=[gold],
                        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, result]
                    )
                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

1478
        elif self.OUTPUT_TYPE == "generate_until":
1479
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1480
            result = results[0]
1481
            if self.config.doc_to_choice is not None:
1482
1483
1484
1485
1486
1487
1488
                try:
                    # If you set doc_to_choice,
                    # it assumes that doc_to_target returns a number.
                    choices = self.doc_to_choice(doc)
                    gold = choices[gold]
                except TypeError:
                    gold = gold
1489
1490
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1491
                gold = list(gold)
Chris's avatar
Chris committed
1492
1493
1494
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1495

lintangsutawika's avatar
lintangsutawika committed
1496
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1497
1498
1499
1500
1501
                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
1502
1503
1504
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        gold = [gold]
1505
1506
1507
1508
1509
1510
1511
1512
                    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
1513
                    else:
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
                        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
1535
                else:
1536
                    try:
JessicaOjo's avatar
JessicaOjo committed
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
                        if metric == "exact_match":
                            result_score = self._metric_fn_list[metric](
                                references=[str(gold)],
                                predictions=[str(result)],
                                **self._metric_fn_kwargs[metric],
                            )
                        else:
                            result_score = self._metric_fn_list[metric](
                                references=[gold],
                                predictions=[result],
                                **self._metric_fn_kwargs[metric],
                            )
JessicaOjo's avatar
JessicaOjo committed
1549
                    except TypeError as error:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1550
                        result_score = self._metric_fn_list[metric]([gold, result])
JessicaOjo's avatar
spacing  
JessicaOjo committed
1551
1552
1553
                    if isinstance(result_score, dict):
                        # TODO: this handles the case where HF evaluate returns a dict.
                        result_score = result_score[metric]
1554
                result_dict[metric] = result_score
1555
        else:
lintangsutawika's avatar
lintangsutawika committed
1556
1557
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1558
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1559
            )
1560
1561
1562

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1563
    def aggregation(self) -> dict:
1564
1565
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1566
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1567
        return self._higher_is_better
1568

Baber Abbasi's avatar
Baber Abbasi committed
1569
1570
1571
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1572
1573
1574
1575
1576
1577
1578
1579
1580
    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)})"
        )

1581
1582

class MultipleChoiceTask(Task):
1583
    OUTPUT_TYPE = "loglikelihood"
1584

baberabb's avatar
baberabb committed
1585
    def doc_to_target(self, doc: dict) -> str:
1586
1587
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1588
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1589
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1590
1591
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1592
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1593
                doc=doc,
1594
                arguments=(ctx, " {}".format(choice)),
1595
                idx=i,
1596
1597
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1598
1599
            for i, choice in enumerate(doc["choices"])
        ]
1600

1601
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1602
1603
1604
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
1615
        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
1616
    def higher_is_better(self) -> dict:
1617
1618
1619
1620
1621
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1622
    def aggregation(self) -> dict:
1623
1624
1625
1626
1627
1628
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1629
class PerplexityTask(Task):
1630
1631
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1632
    def has_training_docs(self) -> bool:
1633
1634
        return False

baberabb's avatar
baberabb committed
1635
    def fewshot_examples(self, k: int, rnd) -> List:
1636
1637
1638
1639
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1640
1641
        return []

baberabb's avatar
baberabb committed
1642
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1643
1644
1645
1646
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1647
1648
1649

        return ""

baberabb's avatar
baberabb committed
1650
    def higher_is_better(self) -> dict:
1651
1652
1653
1654
1655
1656
1657
1658
1659
        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
1660
    def doc_to_text(self, doc) -> str:
1661
1662
1663
1664
1665
        return ""

    def doc_to_target(self, doc):
        return doc

1666
1667
1668
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1669

lintangsutawika's avatar
lintangsutawika committed
1670
1671
1672
1673
1674
1675
1676
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1677

1678
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1679
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1680
1681
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1682
1683
1684
1685
1686
1687
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1688
    def aggregation(self) -> dict:
1689
1690
1691
1692
1693
1694
1695
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1696
    def count_bytes(cls, doc) -> int:
1697
1698
1699
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1700
    def count_words(cls, doc) -> int:
1701
1702
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))