task.py 68.2 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
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
59
    tag: Optional[Union[str, list]] = None
60
61
62
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
63
64
65
66
67
68
    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
69
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
70
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
71
    )
72
73
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
74
75
76
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
77
    doc_to_image: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
78
    unsafe_code: bool = False
79
80
81
    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
Baber Abbasi's avatar
Baber Abbasi committed
96
    gen_prefix: Optional[str] = None
97
98
99
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
100

Ethan Smith's avatar
Ethan Smith committed
101
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
102
        if self.generation_kwargs is not None:
103
            if self.output_type != "generate_until":
104
                eval_logger.warning(
105
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
106
107
108
109
110
111
112
113
                )

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

            if "until" not in self.generation_kwargs:
114
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
115
        else:
116
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
119
120
121
122
123
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
124
125
                    "do_sample": False,
                }
126

127
128
129
    def __getitem__(self, item):
        return getattr(self, item)

130
131
132
    def __setitem__(self, item, value):
        return setattr(self, item, value)

133
    def to_dict(self, keep_callable: bool = False) -> dict:
134
135
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
136
        Used for dumping results alongside full task configuration
137

haileyschoelkopf's avatar
haileyschoelkopf committed
138
139
140
141
142
143
144
145
146
147
        :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)
148
149
150
151
152
153
154
155
156
157
            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
158
        return cfg_dict
159

160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
    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)

176
177
178
179
180
181
182
183
184
185
186

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)
    """

187
    VERSION: Optional[Union[int, str]] = None
188

189
190
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
191
    DATASET_PATH: Optional[str] = None
192
193

    # The name of a subset within `DATASET_PATH`.
194
    DATASET_NAME: Optional[str] = None
195

196
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
197

198
199
    def __init__(
        self,
200
201
202
203
        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
204
    ) -> None:
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
        """
        :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)
227
228
229
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
230

231
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
232

lintangsutawika's avatar
lintangsutawika committed
233
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
234
235
236
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
237

238
239
240
241
242
243
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
        """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.
        """
268
269
270
271
272
273
274
        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,
        )
275

276
    @property
277
    def config(self) -> TaskConfig:
278
279
280
        """Returns the TaskConfig associated with this class."""
        return self._config

281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
    @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

296
    def training_docs(self) -> Iterable:
297
298
299
300
301
302
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

303
    def validation_docs(self) -> Iterable:
304
305
306
307
308
309
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

310
    def test_docs(self) -> Iterable:
311
312
313
314
315
316
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

317
    def fewshot_docs(self) -> Iterable:
318
319
320
321
322
323
324
325
326
        """
        :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
327
            eval_logger.warning(
328
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
329
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
330
            )
331
332
            return self.test_docs()

333
    def _process_doc(self, doc: dict) -> dict:
334
335
336
337
338
339
340
341
342
        """
        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
343

344
    @property
345
    def instances(self) -> List[Instance]:
346
347
348
349
350
351
352
353
354
355
356
        """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)

357
358
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
359
360
361
362
363
364
365
366
367
368
369
            "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

370
371
372
373
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
374
375
376
    def doc_to_prefix(self, doc):
        return ""

377
378
    def build_all_requests(
        self,
379
        *,
380
381
382
383
384
385
386
387
388
389
        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 = "",
390
    ) -> None:
391
        """Build a set of Instances for a task, and store them in task.instances"""
392
393
394
395

        # used with caching
        og_limit = limit

396
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
397
398
399
400
401
402
403
        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 ""
        )
404
        cache_key += f"-tokenizer{tokenizer_name}"
405

Baber Abbasi's avatar
Baber Abbasi committed
406
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
407
408
409
410
411
412
413
414
415
416
417
418
419

        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
420
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
421

422
        instances = []
423
424
425
426
427
428
429
430
431
432

        # 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(
433
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
434
435
436
437
438
439
440
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
441
        ):
442
            # sample fewshot context #TODO: need to offset doc_id by rank now!
443
            fewshot_ctx = self.fewshot_context(
444
                doc,
445
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
446
447
448
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
449
                chat_template,
Baber Abbasi's avatar
Baber Abbasi committed
450
                gen_prefix=self.doc_to_prefix(doc),
451
            )
452

453
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
454
455
456
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
457
                metadata=(self.config["task"], doc_id, self.config.repeats),
458
                apply_chat_template=apply_chat_template,
459
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
460
            )
461
462
463
464

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

465
466
467
468
469
470
471
472
473
474
475
476
477
            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
478

479
480
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
481

482
483
484
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
    @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
501
            The number of times each instance in a dataset is inferred on. Defaults to 1,
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
            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

537
538
539
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
540
541
542
543
544
545
546
547
548
549
    @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))

550
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
551
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
552
553
554
555
556
557
558
        """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
559
560
561
562
563
        :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.
564
565
566
        :returns: str
            The fewshot context.
        """
567
        if rnd is None:
568
569
570
571
572
573
            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
574

575
        description = description if description else ""
576
577

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
578
            labeled_examples = ""
579
        else:
lintangsutawika's avatar
lintangsutawika committed
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
            # 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
604
            )
605
606

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
607
        return description + labeled_examples + example
608

609
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
610
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
611
612
        if hasattr(self, "_filters"):
            for f in self._filters:
613
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
614
615
616
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
617

baberabb's avatar
baberabb committed
618
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
619
        """Returns the config as a dictionary."""
620
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
621
        # (num_fewshot)
622
        return self.config.to_dict()
623

Baber Abbasi's avatar
Baber Abbasi committed
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
659
660
661
662
663
    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)

664
665
666
667
668
    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

669
670
671
672
673
674
675
    @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:
676
677
678
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
679
680
681
682
683
684
685
686
687
688
689
690
691

    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

692
693

class ConfigurableTask(Task):
694
    VERSION = "Yaml"
695
    OUTPUT_TYPE = None
696
    CONFIG = None
697
698

    def __init__(
699
700
701
702
703
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
704
    ) -> None:  # TODO no super() call here
705
        # Get pre-configured attributes
706
        self._config = self.CONFIG
707

708
        # Use new configurations if there was no preconfiguration
709
        if self.config is None:
710
            self._config = TaskConfig(**config)
711
712
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
713
            if config is not None:
714
                self._config.__dict__.update(config)
715

716
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
717
718
719
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
720

721
722
723
724
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

725
        if self.config.output_type is not None:
726
727
728
729
            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)}'"
                )
730
            self.OUTPUT_TYPE = self.config.output_type
731

732
733
734
735
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
736
737
738
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

739
740
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
741

742
743
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
744

745
746
747
748
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
749

750
        if self.config.metric_list is None:
751
            # TODO: handle this in TaskConfig.__post_init__ ?
752
753
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

754
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
755
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
756
                self._metric_fn_kwargs[metric_name] = {}
757
758
759
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
760
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
761
        else:
762
            for metric_config in self.config.metric_list:
763
764
765
766
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
767
768
769
770
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
771
772
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
773
                }
Chris's avatar
Chris committed
774
775
776
777
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
778

779
                if self.config.process_results is not None:
780
781
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
782
783
784
785
786
787
                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
788
789
790
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
791
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
792

793
                if "aggregation" in metric_config:
794
                    agg_name = metric_config["aggregation"]
795
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
796
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
797
                    elif callable(agg_name):  # noqa: E721
798
799
800
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
801
                else:
802
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
803
                    metric_agg = get_metric_aggregation(metric_name)
804
                    eval_logger.warning(
805
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
806
807
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
808
                    )
809
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
810

811
812
813
814
815
816
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
817
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
818
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
819
                        f"higher_is_better={is_higher_better(metric_name)}"
820
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
821
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
822

823
        self.download(self.config.dataset_kwargs)
824
825
826
        self._training_docs = None
        self._fewshot_docs = None

827
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
828
            self._filters = []
829
            for filter_config in self.config.filter_list:
830
831
832
833
834
835
836
837
838
                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
839
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
840
        else:
841
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
842

843
844
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
845
            self.prompt = get_prompt(
846
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
847
            )
848
849
850
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
851
        if self.fewshot_docs() is not None:
852
853
854
855
            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
856
857
858
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
            )
            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)}"
                )
875

876
        self.task_docs = self.eval_docs
877

878
        # Test One Doc
879
        self.features = list(self.task_docs.features.keys())
880
881
        self.multiple_input = 0
        self.multiple_target = 0
882
        test_doc = self.task_docs[0]
883
        test_text = self.doc_to_text(test_doc)
884
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
885

886
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
887
            test_choice = self.doc_to_choice(test_doc)
888
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
889
                eval_logger.error("doc_to_choice must return list")
890
891
            else:
                num_choice = len(test_choice)
892

893
            if isinstance(test_text, int):
894
                self.multiple_input = num_choice
895
896
        else:
            test_choice = None
897

898
        if isinstance(test_target, list):
899
            self.multiple_target = len(test_target)
900
        else:
901
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
902
                test_target = test_choice[test_target]
903
            else:
lintangsutawika's avatar
lintangsutawika committed
904
                test_target = str(test_target)
905

906
907
908
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
909
            check_choices = [test_target]
910
911
912
913
        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 = (
914
915
                    True
                    if self.config.target_delimiter.rstrip()
916
                    != self.config.target_delimiter
917
                    else False
918
                )
919

920
                if delimiter_has_whitespace and choice_has_whitespace:
921
922
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
923
924
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
925
                    eval_logger.debug(
926
                        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'
927
928
                    )

929
    def download(self, dataset_kwargs: Optional[Dict[str, Any]] = None) -> None:
930
931
932
933
934
935
        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
936
    def has_training_docs(self) -> bool:
937
        if self.config.training_split is not None:
938
939
940
941
            return True
        else:
            return False

baberabb's avatar
baberabb committed
942
    def has_validation_docs(self) -> bool:
943
        if self.config.validation_split is not None:
944
945
946
947
            return True
        else:
            return False

baberabb's avatar
baberabb committed
948
    def has_test_docs(self) -> bool:
949
        if self.config.test_split is not None:
950
951
952
953
            return True
        else:
            return False

baberabb's avatar
baberabb committed
954
    def training_docs(self) -> datasets.Dataset:
955
        if self.has_training_docs():
956
957
958
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
959
                )
960
            return self.dataset[self.config.training_split]
961

baberabb's avatar
baberabb committed
962
    def validation_docs(self) -> datasets.Dataset:
963
        if self.has_validation_docs():
964
965
966
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
967
                )
968
            return self.dataset[self.config.validation_split]
969

baberabb's avatar
baberabb committed
970
    def test_docs(self) -> datasets.Dataset:
971
        if self.has_test_docs():
972
973
974
            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]
975

976
    def fewshot_docs(self):
977
        if self.config.fewshot_split is not None:
978
979
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
980
            return self.dataset[self.config.fewshot_split]
981
982
983
984
985
986
987
988
989
990
991
992
        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."
                )
993
        else:
994
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
995
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
996
                    f"[Task: {self.config.task}] "
997
998
999
1000
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1001

KonradSzafer's avatar
KonradSzafer committed
1002
1003
1004
1005
1006
    @staticmethod
    def append_target_question(
        labeled_examples: List[Dict[str, str]],
        question: str,
        fewshot_as_multiturn: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
1007
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    ) -> 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})
Baber Abbasi's avatar
Baber Abbasi committed
1023
1024
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1025

lintangsutawika's avatar
lintangsutawika committed
1026
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1027
1028
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1029
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1030
1031
1032
1033
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1034
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1035
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1036
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1037
1038
1039
1040
1041
1042
1043
        """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
1044
1045
1046
1047
1048
1049
        :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.
1050
1051
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
lintangsutawika's avatar
lintangsutawika committed
1052
1053
1054
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1055
1056
1057
1058
1059
1060
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1061
1062
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1063

KonradSzafer's avatar
KonradSzafer committed
1064
1065
1066
1067
1068
1069
1070
1071
1072
        # 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
1073
        else:
KonradSzafer's avatar
KonradSzafer committed
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
            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(
Baber Abbasi's avatar
Baber Abbasi committed
1087
1088
1089
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1090
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1091
1092
1093
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1094
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1095
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1096
                )
lintangsutawika's avatar
lintangsutawika committed
1097
1098

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1099
1100
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1101
                # TODO: append prefill?
1102
1103
                if not labeled_examples:
                    return ""
1104
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1105
1106
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1107
1108
1109
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1110
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1111
1112
1113
1114
1115
1116
1117
                )
            # 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)
Baber Abbasi's avatar
Baber Abbasi committed
1118
1119
1120
1121
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1122
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1123
1124
1125
1126
1127
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1128
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1129
1130
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1131
1132
1133
1134
1135
1136
                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(
Baber Abbasi's avatar
Baber Abbasi committed
1137
1138
1139
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1140
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1141
1142
1143
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1144
1145
1146
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1147
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1148
1149
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1150
1151
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1152
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1153
            )
1154
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1155
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1156
1157
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1158
1159
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1160
1161
            if self.multiple_input:
                return labeled_examples
1162
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1163
                return labeled_examples + example + prefix
1164
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1165
                return [labeled_examples + ex + prefix for ex in example]
1166
1167
1168
            elif isinstance(example, int):
                if self.config.doc_to_choice is not None:
                    choices = self.doc_to_choice(doc)
Baber Abbasi's avatar
Baber Abbasi committed
1169
                    return labeled_examples + choices[example] + prefix
1170
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1171
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1172

Baber Abbasi's avatar
Baber Abbasi committed
1173
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1174
        """Iterates over FilterEnsembles and applies them to instances"""
1175
1176
        if hasattr(self, "_filters"):
            for f in self._filters:
1177
                f.apply(self._instances)
1178
1179
1180
1181
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1182
    def should_decontaminate(self):
1183
        return self.config.should_decontaminate
1184

Baber Abbasi's avatar
Baber Abbasi committed
1185
    def doc_to_decontamination_query(self, doc: dict):
1186
        if self.config.should_decontaminate:
1187
1188
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1189
            else:
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
                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
                        )
                    )
1201

1202
    def _process_doc(self, doc: dict) -> dict:
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
        """
        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

Yu Shi Jie's avatar
Yu Shi Jie committed
1213
    def doc_to_text(self, doc, doc_to_text=None):
1214
1215
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1216
1217
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1218
        else:
1219
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1220

1221
        if isinstance(doc_to_text, int):
1222
            return doc_to_text
1223
        elif isinstance(doc_to_text, str):
1224
            if doc_to_text in self.features:
1225
                # if self.config.doc_to_choice is not None:
1226
1227
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1228
1229
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1230
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1231
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1232
1233
1234
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1235
        elif callable(doc_to_text):
1236
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1237
        # Used when applying a Promptsource template
1238
        elif hasattr(doc_to_text, "apply"):
1239
1240
1241
1242
1243
            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")
1244
                return self.config.fewshot_delimiter
1245
        else:
1246
            print(type(doc_to_text))
1247
            raise TypeError
1248

Yu Shi Jie's avatar
Yu Shi Jie committed
1249
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1250
1251
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1252
1253
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1254
        else:
1255
            doc_to_target = self.config.doc_to_target
1256

1257
        if isinstance(doc_to_target, int):
1258
            return doc_to_target
1259
        elif isinstance(doc_to_target, str):
1260
            if doc_to_target in self.features:
1261
                # if self.config.doc_to_choice is not None:
1262
1263
1264
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1265
            else:
lintangsutawika's avatar
lintangsutawika committed
1266
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1267
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1268
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1269
1270
1271
1272
1273
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1274
1275
1276
1277
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1278
1279
                else:
                    return target_string
1280
        elif isinstance(doc_to_target, list):
1281
            return doc_to_target
1282
        elif callable(doc_to_target):
1283
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1284
        # Used when applying a Promptsource template
1285
        elif hasattr(doc_to_target, "apply"):
1286
            applied_prompt = doc_to_target.apply(doc)
1287
1288
1289
1290
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1291
                return self.config.fewshot_delimiter
1292
1293
        else:
            raise TypeError
1294

Yu Shi Jie's avatar
Yu Shi Jie committed
1295
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1296
1297
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1298
1299
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1300
        elif self.config.doc_to_choice is None:
1301
1302
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1303
            doc_to_choice = self.config.doc_to_choice
1304

1305
        if isinstance(doc_to_choice, str):
1306
1307
1308
1309
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1310
        elif isinstance(doc_to_choice, list):
1311
            return doc_to_choice
1312
        elif isinstance(doc_to_choice, dict):
1313
1314
1315
1316
1317
1318
1319
            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
1320

1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list]:
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

        if isinstance(doc_to_image, list):
            image_feature = [
                self.doc_to_image(doc, feature) for feature in doc_to_image
            ]
            return [feature for feature in image_feature if feature is not None]
        elif isinstance(doc_to_image, str):
            if doc_to_image in self.features:
                return doc[doc_to_image]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_image, doc))
        elif callable(doc_to_image):
            return doc_to_image(doc)
        else:
            return None

Baber Abbasi's avatar
Baber Abbasi committed
1344
1345
1346
1347
1348
1349
1350
1351
    def doc_to_prefix(self, doc):
        if (gen_prefix := self.config.gen_prefix) is not None:
            if gen_prefix in self.features:
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1352
1353
1354
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1355
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1356
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1357

1358
1359
        aux_arguments = None

1360
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1361
            arguments = (ctx, self.doc_to_target(doc))
1362
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1363
            arguments = (self.doc_to_target(doc),)
1364
        elif self.OUTPUT_TYPE == "multiple_choice":
1365
            choices = self.doc_to_choice(doc)
1366
            target_delimiter = self.config.target_delimiter
1367
1368
            if apply_chat_template:
                target_delimiter = ""
1369
1370
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1371
                # apply chat_template to choices if apply_chat_template
1372
                cont = self.doc_to_target(doc)
1373

1374
                arguments = [
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1385
                ]
1386
            else:
1387
                # Otherwise they are placed in the continuation
1388
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1389

1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # 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.
                aux_arguments = [("", f"{choice}") for choice in choices]

                arguments.extend(aux_arguments)

        elif self.OUTPUT_TYPE == "generate_until":
            arguments = (ctx, deepcopy(self.config.generation_kwargs))

        multimodal_arg = {}
        if (
            self.config.doc_to_image
        ):  # TODO: ensure that non-multimodal tasks aren't getting visual args
            multimodal_arg = {
                **multimodal_arg,
                **{"visual": self.doc_to_image(doc)},
            }

        if bool(multimodal_arg):
            if isinstance(arguments, list):
                arguments = [arg + (multimodal_arg,) for arg in arguments]
            else:
                arguments = arguments + (multimodal_arg,)

        if self.OUTPUT_TYPE == "multiple_choice":
1421
            request_list = [
1422
1423
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1424
                    doc=doc,
1425
                    arguments=arg,
1426
                    idx=i,
1427
1428
                    **kwargs,
                )
1429
                for i, arg in enumerate(arguments)
1430
            ]
1431
1432

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1433

lintangsutawika's avatar
lintangsutawika committed
1434
        return Instance(
1435
1436
1437
1438
1439
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1440
        )
1441
1442

    def process_results(self, doc, results):
1443
1444
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1445

1446
        result_dict = {}
1447
        use_metric = list(self._metric_fn_list.keys())
1448
1449
1450
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1451
1452
1453
1454
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1455
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1456
            (loglikelihood,) = results
1457
1458
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1459
            return {
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
                **(
                    {"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
1475
            }
1476
        elif self.OUTPUT_TYPE == "multiple_choice":
1477
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1478

1479
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1480
            choices = self.doc_to_choice(doc)
1481
1482
            completion_len = np.array([float(len(i)) for i in choices])

1483
1484
            if (
                2 * len(choices) == len(lls)
1485
                and "acc_mutual_info" in self._metric_fn_list.keys()
1486
1487
1488
1489
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1490
1491
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1492
1493
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1494

1495
1496
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1497

1498
1499
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1500
            else:
1501
                gold = self.doc_to_target(doc)
1502
1503

            gold_index_error = False
1504
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1505
1506
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1507
1508
                    gold_index_error = True
            else:
1509
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1510
                    gold = gold if gold < len(choices) else -100
1511
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1512
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1513

Lintang Sutawika's avatar
Lintang Sutawika committed
1514
                if gold == -100:
1515
1516
1517
1518
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1519
                    f"Label index was not in within range of available choices,"
1520
1521
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1522

1523
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1524
1525
                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
1526
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1527
1528
1529
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1530
                # 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
1531
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1532

Lintang Sutawika's avatar
Lintang Sutawika committed
1533
1534
1535
1536
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1537
            result_dict = {
1538
                **({"acc": acc} if "acc" in use_metric else {}),
1539
1540
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1541
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1542
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1543
1544
1545
1546
1547
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1548
1549
            }

1550
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1551
1552
1553
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1554
1555
1556
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1557
        elif self.OUTPUT_TYPE == "generate_until":
1558
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1559
            result = results[0]
1560
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1561
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1562
                # it assumes that doc_to_target returns a number.
1563
1564
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1565
1566
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1567
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1568
1569
1570
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
1571
            ):
Chris's avatar
Chris committed
1572
1573
                # cast gold to the same type as result
                gold = type(result)(gold)
1574

lintangsutawika's avatar
lintangsutawika committed
1575
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1576
1577
1578
1579
1580
                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
1581
1582
1583
1584
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1585
1586
1587
1588
1589
1590
1591
1592
                    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
1593
                    else:
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
1612
1613
1614
                        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
1615
                else:
1616
                    try:
1617
                        result_score = self._metric_fn_list[metric](
1618
1619
                            references=[gold],
                            predictions=[result],
1620
                            **self._metric_fn_kwargs[metric],
1621
                        )
1622
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1623
                        result_score = self._metric_fn_list[metric]([gold, result])
1624
1625
                    if isinstance(result_score, dict):
                        # TODO: this handles the case where HF evaluate returns a dict.
Hojin Lee's avatar
Hojin Lee committed
1626
1627
1628
1629
                        # This allows for multiple metrics to be returned from the same function
                        for k, v in result_score.items():
                            result_dict[k] = v
                        return result_dict
1630
                result_dict[metric] = result_score
1631
        else:
lintangsutawika's avatar
lintangsutawika committed
1632
1633
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1634
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1635
            )
1636
1637
1638

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1639
    def aggregation(self) -> dict:
1640
1641
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1642
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1643
        return self._higher_is_better
1644

Baber Abbasi's avatar
Baber Abbasi committed
1645
1646
1647
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1648
1649
1650
1651
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1652
1653
1654
1655
1656
    def __repr__(self):
        return (
            f"ConfigurableTask(task_name={getattr(self.config, 'task', None)},"
            f"output_type={self.OUTPUT_TYPE},"
            f"num_fewshot={getattr(self.config, 'num_fewshot', None)},"
Baber Abbasi's avatar
Baber Abbasi committed
1657
            f"num_samples={len(self.eval_docs)})"
1658
1659
        )

1660
1661

class MultipleChoiceTask(Task):
1662
    OUTPUT_TYPE = "loglikelihood"
1663

baberabb's avatar
baberabb committed
1664
    def doc_to_target(self, doc: dict) -> str:
1665
1666
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1667
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1668
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1669
1670
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1671
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1672
                doc=doc,
1673
                arguments=(ctx, " {}".format(choice)),
1674
                idx=i,
1675
1676
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1677
1678
            for i, choice in enumerate(doc["choices"])
        ]
1679

1680
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1681
1682
1683
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
        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
1695
    def higher_is_better(self) -> dict:
1696
1697
1698
1699
1700
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1701
    def aggregation(self) -> dict:
1702
1703
1704
1705
1706
1707
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1708
class PerplexityTask(Task):
1709
1710
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1711
    def has_training_docs(self) -> bool:
1712
1713
        return False

baberabb's avatar
baberabb committed
1714
    def fewshot_examples(self, k: int, rnd) -> List:
1715
1716
1717
1718
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1719
1720
        return []

baberabb's avatar
baberabb committed
1721
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1722
1723
1724
1725
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1726
1727
1728

        return ""

baberabb's avatar
baberabb committed
1729
    def higher_is_better(self) -> dict:
1730
1731
1732
1733
1734
1735
1736
1737
1738
        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
1739
    def doc_to_text(self, doc) -> str:
1740
1741
1742
1743
1744
        return ""

    def doc_to_target(self, doc):
        return doc

1745
1746
1747
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1748

lintangsutawika's avatar
lintangsutawika committed
1749
1750
1751
1752
1753
1754
1755
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1756

1757
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1758
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1759
1760
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1761
1762
1763
1764
1765
1766
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1767
    def aggregation(self) -> dict:
1768
1769
1770
1771
1772
1773
1774
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1775
    def count_bytes(cls, doc) -> int:
1776
1777
1778
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1779
    def count_words(cls, doc) -> int:
1780
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1781
        return len(re.split(r"\s+", doc))