task.py 62.5 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
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}"
Konrad's avatar
Konrad committed
387
388
389
390
        cache_key += "-chat_template" if apply_chat_template else ""
        cache_key += "-fewshot_as_multiturn" if fewshot_as_multiturn else ""
        if lm is not None and hasattr(lm, "tokenizer"):
            cache_key += f"-{lm.tokenizer.name_or_path.replace('/', '__')}"
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405

        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
406
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
407

408
        instances = []
409
410
411
412
413
414
415
416
417
418

        # 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(
419
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
420
421
422
423
424
425
426
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
427
        ):
428
            # sample fewshot context #TODO: need to offset doc_id by rank now!
429
            fewshot_ctx = self.fewshot_context(
430
                doc,
431
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
Konrad's avatar
Konrad committed
432
                system_instruction,
Konrad's avatar
Konrad committed
433
                apply_chat_template,
Konrad's avatar
Konrad committed
434
                fewshot_as_multiturn,
Konrad's avatar
Konrad committed
435
                lm,
436
            )
437

438
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
439
440
441
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
442
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
443
            )
444
445
446
447

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

448
449
450
451
452
453
454
455
456
457
458
459
460
            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
461

462
463
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
464

465
466
467
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
    @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
484
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
516
517
518
519
            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

520
521
522
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
523
524
525
526
527
528
529
530
531
532
    @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))

533
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
534
    def fewshot_context(
535
536
537
        self,
        doc,
        num_fewshot,
538
        rnd=None,
539
        description=None,
lintangsutawika's avatar
lintangsutawika committed
540
    ):
541
542
543
544
545
546
547
        """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
548
549
550
551
552
        :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.
553
554
555
        :returns: str
            The fewshot context.
        """
556
        if rnd is None:
557
558
559
560
561
562
            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
563

564
        description = description if description else ""
565
566

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
567
            labeled_examples = ""
568
        else:
lintangsutawika's avatar
lintangsutawika committed
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
            # 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
593
            )
594
595

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
596
        return description + labeled_examples + example
597

598
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
599
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
600
601
        if hasattr(self, "_filters"):
            for f in self._filters:
602
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
603
604
605
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
606

baberabb's avatar
baberabb committed
607
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
608
        """Returns the config as a dictionary."""
609
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
610
        # (num_fewshot)
611
        return self.config.to_dict()
612

Baber Abbasi's avatar
Baber Abbasi committed
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
649
650
651
652
    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)

653
654
655
656
657
    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

658
659
660
661
662
663
664
    @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:
665
666
667
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
668
669
670
671
672
673
674
675
676
677
678
679
680

    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

681
682

class ConfigurableTask(Task):
683
    VERSION = "Yaml"
684
    OUTPUT_TYPE = None
685
    CONFIG = None
686
687

    def __init__(
688
689
690
691
692
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
693
    ) -> None:  # TODO no super() call here
694
        # Get pre-configured attributes
695
        self._config = self.CONFIG
696

697
        # Use new configurations if there was no preconfiguration
698
        if self.config is None:
699
            self._config = TaskConfig(**config)
700
701
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
702
            if config is not None:
703
                self._config.__dict__.update(config)
704

705
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
706
707
708
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
709

710
711
712
713
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

714
        if self.config.output_type is not None:
715
716
717
718
            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)}'"
                )
719
            self.OUTPUT_TYPE = self.config.output_type
720

721
722
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
723

724
725
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
726

727
728
729
730
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
731

732
        if self.config.metric_list is None:
733
            # TODO: handle this in TaskConfig.__post_init__ ?
734
735
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

736
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
737
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
738
                self._metric_fn_kwargs[metric_name] = {}
739
740
741
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
742
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
743
        else:
744
            for metric_config in self.config.metric_list:
745
746
747
748
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
749
750
751
752
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
753
754
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
755
                }
Chris's avatar
Chris committed
756
757
758
759
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
760

761
                if self.config.process_results is not None:
762
763
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
764
765
766
767
768
769
                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
770
771
772
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
773
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
774

775
                if "aggregation" in metric_config:
776
                    agg_name = metric_config["aggregation"]
777
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
778
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
779
                    elif callable(agg_name):  # noqa: E721
780
781
782
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
783
                else:
784
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
785
                    metric_agg = get_metric_aggregation(metric_name)
786
                    eval_logger.warning(
787
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
788
789
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
790
                    )
791
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
792

793
794
795
796
797
798
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
799
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
800
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
801
                        f"higher_is_better={is_higher_better(metric_name)}"
802
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
803
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
804

805
        self.download(self.config.dataset_kwargs)
806
807
808
        self._training_docs = None
        self._fewshot_docs = None

809
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
810
            self._filters = []
811
            for filter_config in self.config.filter_list:
812
813
814
815
816
817
818
819
820
                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
821
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
822
        else:
823
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
824

825
826
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
827
            self.prompt = get_prompt(
828
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
829
            )
830
831
832
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
833
        if self.fewshot_docs() is not None:
834
835
836
837
            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
838
839
840
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
            )
            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)}"
                )
857

858
        self.task_docs = self.eval_docs
859

860
        # Test One Doc
861
        self.features = list(self.task_docs.features.keys())
862
863
        self.multiple_input = 0
        self.multiple_target = 0
864
        test_doc = self.task_docs[0]
865
        test_text = self.doc_to_text(test_doc)
866
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
867

868
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
869
            test_choice = self.doc_to_choice(test_doc)
870
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
871
                eval_logger.error("doc_to_choice must return list")
872
873
            else:
                num_choice = len(test_choice)
874

875
            if isinstance(test_text, int):
876
                self.multiple_input = num_choice
877
878
        else:
            test_choice = None
879

880
        if isinstance(test_target, list):
881
            self.multiple_target = len(test_target)
882
        else:
883
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
884
                test_target = test_choice[test_target]
885
            else:
lintangsutawika's avatar
lintangsutawika committed
886
                test_target = str(test_target)
887

888
889
890
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
891
            check_choices = [test_target]
892
893
894
895
        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 = (
896
897
                    True
                    if self.config.target_delimiter.rstrip()
898
                    != self.config.target_delimiter
899
                    else False
900
                )
901

902
                if delimiter_has_whitespace and choice_has_whitespace:
903
904
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
905
906
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
907
                    eval_logger.debug(
908
                        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'
909
910
                    )

911
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
912
913
914
915
916
917
        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
918
    def has_training_docs(self) -> bool:
919
        if self.config.training_split is not None:
920
921
922
923
            return True
        else:
            return False

baberabb's avatar
baberabb committed
924
    def has_validation_docs(self) -> bool:
925
        if self.config.validation_split is not None:
926
927
928
929
            return True
        else:
            return False

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

baberabb's avatar
baberabb committed
936
    def training_docs(self) -> datasets.Dataset:
937
        if self.has_training_docs():
938
939
940
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
941
                )
942
            return self.dataset[self.config.training_split]
943

baberabb's avatar
baberabb committed
944
    def validation_docs(self) -> datasets.Dataset:
945
        if self.has_validation_docs():
946
947
948
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
949
                )
950
            return self.dataset[self.config.validation_split]
951

baberabb's avatar
baberabb committed
952
    def test_docs(self) -> datasets.Dataset:
953
        if self.has_test_docs():
954
955
956
            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]
957

958
    def fewshot_docs(self):
959
        if self.config.fewshot_split is not None:
960
961
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
962
            return self.dataset[self.config.fewshot_split]
963
        else:
964
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
965
                eval_logger.warning(
966
                    f"Task '{self.config.task}': "
967
968
969
970
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
971

Konrad's avatar
Konrad committed
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
    @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
993
    @utils.positional_deprecated
Konrad's avatar
Konrad committed
994
995
996
997
    def fewshot_context(
        self,
        doc: str,
        num_fewshot: int,
Konrad's avatar
Konrad committed
998
        system_instruction: Optional[str] = None,
Konrad's avatar
Konrad committed
999
        apply_chat_template: bool = False,
Konrad's avatar
Konrad committed
1000
        fewshot_as_multiturn: bool = False,
Konrad's avatar
Konrad committed
1001
        lm=None,
Konrad's avatar
Konrad committed
1002
    ) -> str:
lintangsutawika's avatar
lintangsutawika committed
1003
1004
1005
1006
1007
1008
1009
        """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
1010
1011
        :param  system_instruction: str
            System instruction to be applied to the prompt.
Konrad's avatar
Konrad committed
1012
1013
        :param apply_chat_template: bool
            Whether to apply the chat template to the fewshot context.
Konrad's avatar
Konrad committed
1014
1015
        :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
1016
1017
        :param lm:
            Language model with definition of the tokenizer/function to use for applying the chat template.
lintangsutawika's avatar
lintangsutawika committed
1018
1019
1020
        :returns: str
            The fewshot context.
        """
Konrad's avatar
Konrad committed
1021
1022
1023
1024
1025
1026
1027

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

        # get task description
1028
1029
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1030

Konrad's avatar
Konrad committed
1031
        # create system prompt based on the provided system instruction and description
Konrad's avatar
Konrad committed
1032
        if system_instruction is not None and description:
Konrad's avatar
Konrad committed
1033
1034
1035
            system_prompt = (
                f"{system_instruction}{self.sampler.fewshot_delimiter}{description}"
            )
Konrad's avatar
Konrad committed
1036
        elif system_instruction is not None:
Konrad's avatar
Konrad committed
1037
1038
1039
1040
1041
1042
1043
1044
            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
1045
            if apply_chat_template:
Konrad's avatar
Konrad committed
1046
                labeled_examples.append({"role": "system", "content": system_prompt})
Konrad's avatar
Konrad committed
1047
            else:
Konrad's avatar
Konrad committed
1048
1049
1050
1051
                labeled_examples = system_prompt

        # if few-shot - append examples after the system prompt
        if num_fewshot > 0:
Konrad's avatar
Konrad committed
1052
            if apply_chat_template:
1053
1054
1055
1056
                labeled_examples.extend(
                    self.sampler.get_chat_context(
                        doc, num_fewshot, fewshot_as_multiturn
                    )
Konrad's avatar
Konrad committed
1057
1058
                )
            else:
Konrad's avatar
Konrad committed
1059
                labeled_examples += self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1060
1061

        example = self.doc_to_text(doc)
Konrad's avatar
Konrad committed
1062
        if apply_chat_template:
Konrad's avatar
Konrad committed
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
            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
1090
            return lm.apply_chat_template(labeled_examples)
1091
        else:
Konrad's avatar
Konrad committed
1092
1093
            if self.multiple_input:
                return labeled_examples
Konrad's avatar
Konrad committed
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
            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
1104

1105
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1106
        """Iterates over FilterEnsembles and applies them to instances"""
1107
1108
        if hasattr(self, "_filters"):
            for f in self._filters:
1109
                f.apply(self._instances)
1110
1111
1112
1113
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1114
    def should_decontaminate(self):
1115
        return self.config.should_decontaminate
1116
1117

    def doc_to_decontamination_query(self, doc):
1118
        if self.config.should_decontaminate:
1119
1120
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1121
            else:
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
                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
                        )
                    )
1133

1134
    def _process_doc(self, doc: dict) -> dict:
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
        """
        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):
1146
1147
        if self.prompt is not None:
            doc_to_text = self.prompt
1148
        else:
1149
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1150

1151
        if isinstance(doc_to_text, int):
1152
            return doc_to_text
1153
        elif isinstance(doc_to_text, str):
1154
            if doc_to_text in self.features:
1155
                # if self.config.doc_to_choice is not None:
1156
1157
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1158
1159
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1160
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1161
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1162
1163
1164
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1165
        elif callable(doc_to_text):
1166
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1167
        # Used when applying a Promptsource template
1168
        elif hasattr(doc_to_text, "apply"):
1169
1170
1171
1172
1173
            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")
1174
                return self.config.fewshot_delimiter
1175
        else:
1176
            print(type(doc_to_text))
1177
            raise TypeError
1178

1179
    def doc_to_target(self, doc: Mapping) -> Union[int, str, list]:
1180
1181
        if self.prompt is not None:
            doc_to_target = self.prompt
1182
        else:
1183
            doc_to_target = self.config.doc_to_target
1184

1185
        if isinstance(doc_to_target, int):
1186
            return doc_to_target
1187
        elif isinstance(doc_to_target, str):
1188
            if doc_to_target in self.features:
1189
                # if self.config.doc_to_choice is not None:
1190
1191
1192
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1193
            else:
lintangsutawika's avatar
lintangsutawika committed
1194
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1195
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1196
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1197
1198
1199
1200
1201
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1202
1203
1204
1205
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1206
1207
                else:
                    return target_string
1208
        elif isinstance(doc_to_target, list):
1209
            return doc_to_target
1210
        elif callable(doc_to_target):
1211
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1212
        # Used when applying a Promptsource template
1213
        elif hasattr(doc_to_target, "apply"):
1214
            applied_prompt = doc_to_target.apply(doc)
1215
1216
1217
1218
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1219
                return self.config.fewshot_delimiter
1220
1221
        else:
            raise TypeError
1222

baberabb's avatar
baberabb committed
1223
    def doc_to_choice(self, doc: Any) -> List[str]:
1224
1225
        if self.prompt is not None:
            doc_to_choice = self.prompt
1226
        elif self.config.doc_to_choice is None:
1227
1228
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1229
            doc_to_choice = self.config.doc_to_choice
1230

1231
        if isinstance(doc_to_choice, str):
1232
1233
1234
1235
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1236
        elif isinstance(doc_to_choice, list):
1237
            return doc_to_choice
1238
        elif isinstance(doc_to_choice, dict):
1239
1240
1241
1242
1243
1244
1245
            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
1246

baberabb's avatar
baberabb committed
1247
1248
1249
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1250
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1251
            arguments = (ctx, self.doc_to_target(doc))
1252
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1253
            arguments = (self.doc_to_target(doc),)
1254
        elif self.OUTPUT_TYPE == "multiple_choice":
1255
            choices = self.doc_to_choice(doc)
1256
            target_delimiter = self.config.target_delimiter
1257
1258
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1259
                cont = self.doc_to_target(doc)
1260
1261
1262
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1263
            else:
1264
                # Otherwise they are placed in the continuation
1265
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1266

1267
            request_list = [
1268
1269
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1270
                    doc=doc,
1271
                    arguments=arg,
1272
                    idx=i,
1273
1274
                    **kwargs,
                )
1275
                for i, arg in enumerate(arguments)
1276
            ]
1277
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1278
            if "acc_mutual_info" in self._metric_fn_list.keys():
1279
1280
1281
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1282
                # here mutual info refers to calculating
1283
1284
1285
1286
1287
1288
                # 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
1289
                            doc=doc,
1290
                            arguments=("", "{}".format(choice)),
1291
1292
1293
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1294
                        for i, choice in enumerate(choices)
1295
1296
1297
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1298

1299
        elif self.OUTPUT_TYPE == "generate_until":
1300
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1301
1302

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1303
1304
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1305
1306

    def process_results(self, doc, results):
1307
1308
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1309

1310
        result_dict = {}
1311
        use_metric = list(self._metric_fn_list.keys())
1312
1313
1314
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1315
1316
1317
1318
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1319
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1320
            (loglikelihood,) = results
1321
1322
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1323
            return {
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
                **(
                    {"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
1339
            }
1340
        elif self.OUTPUT_TYPE == "multiple_choice":
1341
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1342

1343
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1344
            choices = self.doc_to_choice(doc)
1345
1346
            completion_len = np.array([float(len(i)) for i in choices])

1347
1348
            if (
                2 * len(choices) == len(lls)
1349
                and "acc_mutual_info" in self._metric_fn_list.keys()
1350
1351
1352
1353
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1354
1355
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1356
1357
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1358

1359
1360
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1361

1362
1363
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1364
            else:
1365
                gold = self.doc_to_target(doc)
1366
1367

            gold_index_error = False
1368
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1369
1370
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1371
1372
                    gold_index_error = True
            else:
1373
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1374
                    gold = gold if gold < len(choices) else -100
1375
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1376
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1377

Lintang Sutawika's avatar
Lintang Sutawika committed
1378
                if gold == -100:
1379
1380
1381
1382
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1383
                    f"Label index was not in within range of available choices,"
1384
1385
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1386

1387
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1388
1389
                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
1390
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1391
1392
1393
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1394
                # 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
1395
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1396

Lintang Sutawika's avatar
Lintang Sutawika committed
1397
1398
1399
1400
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1401
            result_dict = {
1402
                **({"acc": acc} if "acc" in use_metric else {}),
1403
1404
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1405
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1406
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1407
1408
1409
1410
1411
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1412
1413
            }

1414
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1415
1416
1417
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1418
1419
1420
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1421
        elif self.OUTPUT_TYPE == "generate_until":
1422
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1423
            result = results[0]
1424
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1425
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1426
                # it assumes that doc_to_target returns a number.
1427
1428
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1429
1430
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1431
                gold = list(gold)
Chris's avatar
Chris committed
1432
1433
1434
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1435

lintangsutawika's avatar
lintangsutawika committed
1436
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1437
1438
1439
1440
1441
                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
1442
1443
1444
1445
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1446
1447
1448
1449
1450
1451
1452
1453
                    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
1454
                    else:
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
                        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
1476
                else:
1477
                    try:
1478
                        result_score = self._metric_fn_list[metric](
1479
1480
                            references=[gold],
                            predictions=[result],
1481
                            **self._metric_fn_kwargs[metric],
1482
                        )
1483
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1484
                        result_score = self._metric_fn_list[metric]([gold, result])
1485
1486
1487
1488
                    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
1489
        else:
lintangsutawika's avatar
lintangsutawika committed
1490
1491
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1492
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1493
            )
1494
1495
1496

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1497
    def aggregation(self) -> dict:
1498
1499
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1500
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1501
        return self._higher_is_better
1502

Baber Abbasi's avatar
Baber Abbasi committed
1503
1504
1505
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

1506
1507
1508
1509
1510
1511
1512
1513
1514
    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)})"
        )

1515
1516

class MultipleChoiceTask(Task):
1517
    OUTPUT_TYPE = "loglikelihood"
1518

baberabb's avatar
baberabb committed
1519
    def doc_to_target(self, doc: dict) -> str:
1520
1521
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1522
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1523
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1524
1525
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1526
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1527
                doc=doc,
1528
                arguments=(ctx, " {}".format(choice)),
1529
                idx=i,
1530
1531
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1532
1533
            for i, choice in enumerate(doc["choices"])
        ]
1534

1535
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1536
1537
1538
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
        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
1550
    def higher_is_better(self) -> dict:
1551
1552
1553
1554
1555
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1556
    def aggregation(self) -> dict:
1557
1558
1559
1560
1561
1562
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1563
class PerplexityTask(Task):
1564
1565
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1566
    def has_training_docs(self) -> bool:
1567
1568
        return False

baberabb's avatar
baberabb committed
1569
    def fewshot_examples(self, k: int, rnd) -> List:
1570
1571
1572
1573
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1574
1575
        return []

baberabb's avatar
baberabb committed
1576
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1577
1578
1579
1580
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1581
1582
1583

        return ""

baberabb's avatar
baberabb committed
1584
    def higher_is_better(self) -> dict:
1585
1586
1587
1588
1589
1590
1591
1592
1593
        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
1594
    def doc_to_text(self, doc) -> str:
1595
1596
1597
1598
1599
        return ""

    def doc_to_target(self, doc):
        return doc

1600
1601
1602
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1603

lintangsutawika's avatar
lintangsutawika committed
1604
1605
1606
1607
1608
1609
1610
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1611

1612
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1613
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1614
1615
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1616
1617
1618
1619
1620
1621
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1622
    def aggregation(self) -> dict:
1623
1624
1625
1626
1627
1628
1629
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1630
    def count_bytes(cls, doc) -> int:
1631
1632
1633
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1634
    def count_words(cls, doc) -> int:
1635
1636
        """Downstream tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))