task.py 64.7 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
    group: 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
    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
70
71
72
    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
    doc_to_decontamination_query: Optional[str] = None
95
    gen_prefix: str = None
96
97
98
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
99

Ethan Smith's avatar
Ethan Smith committed
100
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
101
102
103
104
105
106
107
108
109
110
111
112
        if self.group is not None:
            eval_logger.warning(
                "A task YAML file was found to contain a `group` key. Groups which provide aggregate scores over several subtasks now require a separate config file--if not aggregating, you may want to use the `tag` config option instead within your config. Setting `group` within a TaskConfig will be deprecated in v0.4.4. Please see https://github.com/EleutherAI/lm-evaluation-harness/blob/main/docs/task_guide.md for more information."
            )

            if self.tag is None:
                self.tag = self.group
            else:
                raise ValueError(
                    "Got both a `group` and `tag` entry within a TaskConfig. Please use one or the other--`group` values will be deprecated in v0.4.4."
                )

Lintang Sutawika's avatar
Lintang Sutawika committed
113
        if self.generation_kwargs is not None:
114
            if self.output_type != "generate_until":
115
                eval_logger.warning(
116
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
117
118
119
120
121
122
123
124
                )

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

            if "until" not in self.generation_kwargs:
125
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
126
        else:
127
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
128
129
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
130
131
132
133
134
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
135
136
                    "do_sample": False,
                }
137

138
139
140
    def __getitem__(self, item):
        return getattr(self, item)

141
142
143
    def __setitem__(self, item, value):
        return setattr(self, item, value)

144
    def to_dict(self, keep_callable: bool = False) -> dict:
145
146
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
147
        Used for dumping results alongside full task configuration
148

haileyschoelkopf's avatar
haileyschoelkopf committed
149
150
151
152
153
154
155
156
157
158
        :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)
159
160
161
162
163
164
165
166
167
168
            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
169
        return cfg_dict
170

171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
    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)

187
188
189
190
191
192
193
194
195
196
197

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)
    """

198
    VERSION: Optional[Union[int, str]] = None
199

200
201
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
202
    DATASET_PATH: Optional[str] = None
203
204

    # The name of a subset within `DATASET_PATH`.
205
    DATASET_NAME: Optional[str] = None
206

207
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
208

209
210
    def __init__(
        self,
211
212
213
214
        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
215
    ) -> None:
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
        """
        :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)
238
239
240
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
241

242
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
243

lintangsutawika's avatar
lintangsutawika committed
244
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
245
246
247
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
248

249
250
251
252
253
254
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
        """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.
        """
279
280
281
282
283
284
285
        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,
        )
286

287
    @property
288
    def config(self) -> TaskConfig:
289
290
291
        """Returns the TaskConfig associated with this class."""
        return self._config

292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
    @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

307
    def training_docs(self) -> Iterable:
308
309
310
311
312
313
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

314
    def validation_docs(self) -> Iterable:
315
316
317
318
319
320
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

321
    def test_docs(self) -> Iterable:
322
323
324
325
326
327
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

328
    def fewshot_docs(self) -> Iterable:
329
330
331
332
333
334
335
336
337
        """
        :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
338
            eval_logger.warning(
339
                f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
340
                ", using test_docs as fewshot_docs but this is not recommended."
lintangsutawika's avatar
lintangsutawika committed
341
            )
342
343
            return self.test_docs()

344
    def _process_doc(self, doc: dict) -> dict:
345
346
347
348
349
350
351
352
353
        """
        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
354

355
    @property
356
    def instances(self) -> List[Instance]:
357
358
359
360
361
362
363
364
365
366
367
        """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)

368
369
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
370
371
372
373
374
375
376
377
378
379
380
            "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

381
382
    def build_all_requests(
        self,
383
        *,
384
385
386
387
388
389
390
391
392
393
        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 = "",
394
    ) -> None:
395
        """Build a set of Instances for a task, and store them in task.instances"""
396
397
398
399

        # used with caching
        og_limit = limit

400
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
401
402
403
404
405
406
407
        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 ""
        )
408
        cache_key += f"-tokenizer{tokenizer_name}"
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423

        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
424
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
425

426
        instances = []
427
428
429
430
431
432
433
434
435
436

        # 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(
437
            self.doc_iterator(rank=rank, limit=limit, world_size=world_size)
438
439
440
441
442
443
444
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
445
        ):
446
            # sample fewshot context #TODO: need to offset doc_id by rank now!
447
            fewshot_ctx = self.fewshot_context(
448
                doc,
449
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
450
451
452
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
453
                chat_template,
454
            )
455

456
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
457
458
459
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
460
                metadata=(self.config["task"], doc_id, self.config.repeats),
lintangsutawika's avatar
lintangsutawika committed
461
            )
462
463
464
465

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

466
467
468
469
470
471
472
473
474
475
476
477
478
            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
479

480
481
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
482

483
484
485
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
    @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
502
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
537
            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

538
539
540
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
541
542
543
544
545
546
547
548
549
550
    @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))

551
    @utils.positional_deprecated
lintangsutawika's avatar
lintangsutawika committed
552
    def fewshot_context(
553
554
555
        self,
        doc,
        num_fewshot,
556
        rnd=None,
557
        description=None,
lintangsutawika's avatar
lintangsutawika committed
558
    ):
559
560
561
562
563
564
565
        """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
566
567
568
569
570
        :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.
571
572
573
        :returns: str
            The fewshot context.
        """
574
        if rnd is None:
575
576
577
578
579
580
            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
581

582
        description = description if description else ""
583
584

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
585
            labeled_examples = ""
586
        else:
lintangsutawika's avatar
lintangsutawika committed
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
            # 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
611
            )
612
613

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
614
        return description + labeled_examples + example
615

616
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
617
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
618
619
        if hasattr(self, "_filters"):
            for f in self._filters:
620
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
621
622
623
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
624

baberabb's avatar
baberabb committed
625
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
626
        """Returns the config as a dictionary."""
627
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
628
        # (num_fewshot)
629
        return self.config.to_dict()
630

Baber Abbasi's avatar
Baber Abbasi committed
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
664
665
666
667
668
669
670
    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)

671
672
673
674
675
    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

676
677
678
679
680
681
682
    @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:
683
684
685
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
686
687
688
689
690
691
692
693
694
695
696
697
698

    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

699
700

class ConfigurableTask(Task):
701
    VERSION = "Yaml"
702
    OUTPUT_TYPE = None
703
    CONFIG = None
704
705

    def __init__(
706
707
708
709
710
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
711
    ) -> None:  # TODO no super() call here
712
        # Get pre-configured attributes
713
        self._config = self.CONFIG
714

715
        # Use new configurations if there was no preconfiguration
716
        if self.config is None:
717
            self._config = TaskConfig(**config)
718
719
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
720
            if config is not None:
721
                self._config.__dict__.update(config)
722

723
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
724
725
726
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
727

728
729
730
731
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

732
        if self.config.output_type is not None:
733
734
735
736
            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)}'"
                )
737
            self.OUTPUT_TYPE = self.config.output_type
738

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
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
    @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
1023
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1024
1025
    def fewshot_context(
        self,
1026
        doc: Dict[str, str],
KonradSzafer's avatar
KonradSzafer committed
1027
1028
1029
1030
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1031
        chat_template: Optional[Callable] = None,
1032
    ) -> Union[str, List[Dict[str, str]]]:
lintangsutawika's avatar
lintangsutawika committed
1033
1034
1035
1036
1037
1038
1039
        """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
1040
1041
1042
1043
1044
1045
        :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.
1046
1047
        :param chat_template: Callable
            Chat template to be applied to the fewshot context.
lintangsutawika's avatar
lintangsutawika committed
1048
1049
1050
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1051
1052
1053
1054
1055
1056
1057

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

        # get task description
1058
1059
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1060

1061
        ## [System prompt/Description] ##
KonradSzafer's avatar
KonradSzafer committed
1062
1063
1064
1065
1066
1067
1068
1069
1070
        # 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
1071
        else:
KonradSzafer's avatar
KonradSzafer committed
1072
1073
1074
1075
1076
1077
1078
1079
1080
            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

1081
        ## [Few-shot examples] ##
KonradSzafer's avatar
KonradSzafer committed
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
        # if few-shot - append examples after the system prompt
        if num_fewshot > 0:
            if apply_chat_template:
                labeled_examples.extend(
                    self.sampler.get_chat_context(
                        doc, num_fewshot, fewshot_as_multiturn
                    )
                )
            else:
                labeled_examples += self.sampler.get_context(doc, num_fewshot)
lintangsutawika's avatar
lintangsutawika committed
1092

1093
        ## [Prompt] ##
lintangsutawika's avatar
lintangsutawika committed
1094
        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1095
1096
        if apply_chat_template:
            if self.multiple_input:
1097
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
            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)
1109
                    labeled_examples_list.append(chat_template(chat))
KonradSzafer's avatar
KonradSzafer committed
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
                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)
1123
            return chat_template(labeled_examples)
1124
        else:
KonradSzafer's avatar
KonradSzafer committed
1125
1126
            if self.multiple_input:
                return labeled_examples
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
            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
1137

1138
    def apply_filters(self):
Baber Abbasi's avatar
Baber Abbasi committed
1139
        """Iterates over FilterEnsembles and applies them to instances"""
1140
1141
        if hasattr(self, "_filters"):
            for f in self._filters:
1142
                f.apply(self._instances)
1143
1144
1145
1146
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1147
    def should_decontaminate(self):
1148
        return self.config.should_decontaminate
1149
1150

    def doc_to_decontamination_query(self, doc):
1151
        if self.config.should_decontaminate:
1152
1153
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1154
            else:
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
                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
                        )
                    )
1166

1167
    def _process_doc(self, doc: dict) -> dict:
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
        """
        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
1178
    def doc_to_text(self, doc, doc_to_text=None):
1179
1180
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1181
1182
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1183
        else:
1184
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1185

1186
        if isinstance(doc_to_text, int):
1187
            return doc_to_text
1188
        elif isinstance(doc_to_text, str):
1189
            if doc_to_text in self.features:
1190
                # if self.config.doc_to_choice is not None:
1191
1192
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1193
1194
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1195
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1196
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1197
1198
1199
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1200
        elif callable(doc_to_text):
1201
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1202
        # Used when applying a Promptsource template
1203
        elif hasattr(doc_to_text, "apply"):
1204
1205
1206
1207
1208
            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")
1209
                return self.config.fewshot_delimiter
1210
        else:
1211
            print(type(doc_to_text))
1212
            raise TypeError
1213

1214
    def doc_to_target(self, doc: Dict, doc_to_target=None) -> Union[int, str, list]:
1215
1216
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1217
1218
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1219
        else:
1220
            doc_to_target = self.config.doc_to_target
1221

1222
        if isinstance(doc_to_target, int):
1223
            return doc_to_target
1224
        elif isinstance(doc_to_target, str):
1225
            if doc_to_target in self.features:
1226
                # if self.config.doc_to_choice is not None:
1227
1228
1229
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1230
            else:
lintangsutawika's avatar
lintangsutawika committed
1231
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1232
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1233
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1234
1235
1236
1237
1238
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1239
1240
1241
1242
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1243
1244
                else:
                    return target_string
1245
        elif isinstance(doc_to_target, list):
1246
            return doc_to_target
1247
        elif callable(doc_to_target):
1248
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1249
        # Used when applying a Promptsource template
1250
        elif hasattr(doc_to_target, "apply"):
1251
            applied_prompt = doc_to_target.apply(doc)
1252
1253
1254
1255
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1256
                return self.config.fewshot_delimiter
1257
1258
        else:
            raise TypeError
1259

Yu Shi Jie's avatar
Yu Shi Jie committed
1260
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1261
1262
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1263
1264
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1265
        elif self.config.doc_to_choice is None:
1266
1267
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1268
            doc_to_choice = self.config.doc_to_choice
1269

1270
        if isinstance(doc_to_choice, str):
1271
1272
1273
1274
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1275
        elif isinstance(doc_to_choice, list):
1276
            return doc_to_choice
1277
        elif isinstance(doc_to_choice, dict):
1278
1279
1280
1281
1282
1283
1284
            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
1285

baberabb's avatar
baberabb committed
1286
1287
1288
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1289
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1290
            arguments = (ctx, self.doc_to_target(doc))
1291
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1292
            arguments = (self.doc_to_target(doc),)
1293
        elif self.OUTPUT_TYPE == "multiple_choice":
1294
            choices = self.doc_to_choice(doc)
1295
            target_delimiter = self.config.target_delimiter
1296
1297
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1298
                cont = self.doc_to_target(doc)
1299
1300
1301
                arguments = [
                    (ctx + choice, f"{target_delimiter}{cont}") for choice in choices
                ]
1302
            else:
1303
                # Otherwise they are placed in the continuation
1304
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1305

1306
            request_list = [
1307
1308
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1309
                    doc=doc,
1310
                    arguments=arg,
1311
                    idx=i,
1312
1313
                    **kwargs,
                )
1314
                for i, arg in enumerate(arguments)
1315
            ]
1316
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1317
            if "acc_mutual_info" in self._metric_fn_list.keys():
1318
1319
1320
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

lintangsutawika's avatar
lintangsutawika committed
1321
                # here mutual info refers to calculating
1322
1323
1324
1325
1326
1327
                # 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
1328
                            doc=doc,
1329
                            arguments=("", "{}".format(choice)),
1330
1331
1332
                            idx=i,
                            **kwargs,
                        )
lintangsutawika's avatar
lintangsutawika committed
1333
                        for i, choice in enumerate(choices)
1334
1335
1336
                    ]
                )
            return request_list
lintangsutawika's avatar
lintangsutawika committed
1337

1338
        elif self.OUTPUT_TYPE == "generate_until":
1339
            arguments = (ctx, deepcopy(self.config.generation_kwargs))
lintangsutawika's avatar
lintangsutawika committed
1340
1341

        return Instance(
lintangsutawika's avatar
lintangsutawika committed
1342
1343
            request_type=self.OUTPUT_TYPE, doc=doc, arguments=arguments, idx=0, **kwargs
        )
1344
1345

    def process_results(self, doc, results):
1346
1347
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1348

1349
        result_dict = {}
1350
        use_metric = list(self._metric_fn_list.keys())
1351
1352
1353
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1354
1355
1356
1357
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1358
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1359
            (loglikelihood,) = results
1360
1361
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1362
            return {
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
                **(
                    {"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
1378
            }
1379
        elif self.OUTPUT_TYPE == "multiple_choice":
1380
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1381

1382
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1383
            choices = self.doc_to_choice(doc)
1384
1385
            completion_len = np.array([float(len(i)) for i in choices])

1386
1387
            if (
                2 * len(choices) == len(lls)
1388
                and "acc_mutual_info" in self._metric_fn_list.keys()
1389
1390
1391
1392
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1393
1394
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1395
1396
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1397

1398
1399
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1400

1401
1402
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1403
            else:
1404
                gold = self.doc_to_target(doc)
1405
1406

            gold_index_error = False
1407
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1408
1409
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1410
1411
                    gold_index_error = True
            else:
1412
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1413
                    gold = gold if gold < len(choices) else -100
1414
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1415
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1416

Lintang Sutawika's avatar
Lintang Sutawika committed
1417
                if gold == -100:
1418
1419
1420
1421
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1422
                    f"Label index was not in within range of available choices,"
1423
1424
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1425

1426
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1427
1428
                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
1429
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1430
1431
1432
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1433
                # 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
1434
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1435

Lintang Sutawika's avatar
Lintang Sutawika committed
1436
1437
1438
1439
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1440
            result_dict = {
1441
                **({"acc": acc} if "acc" in use_metric else {}),
1442
1443
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1444
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1445
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1446
1447
1448
1449
1450
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1451
1452
            }

1453
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1454
1455
1456
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1457
1458
1459
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1460
        elif self.OUTPUT_TYPE == "generate_until":
1461
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1462
            result = results[0]
1463
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1464
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1465
                # it assumes that doc_to_target returns a number.
1466
1467
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1468
1469
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1470
                gold = list(gold)
Chris's avatar
Chris committed
1471
1472
1473
            elif type(gold) != type(result):
                # cast gold to the same type as result
                gold = type(result)(gold)
1474

lintangsutawika's avatar
lintangsutawika committed
1475
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1476
1477
1478
1479
1480
                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
1481
1482
1483
1484
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1485
1486
1487
1488
1489
1490
1491
1492
                    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
1493
                    else:
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
                        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
1515
                else:
1516
                    try:
1517
                        result_score = self._metric_fn_list[metric](
1518
1519
                            references=[gold],
                            predictions=[result],
1520
                            **self._metric_fn_kwargs[metric],
1521
                        )
1522
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1523
                        result_score = self._metric_fn_list[metric]([gold, result])
1524
1525
1526
1527
                    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
1528
        else:
lintangsutawika's avatar
lintangsutawika committed
1529
1530
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1531
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1532
            )
1533
1534
1535

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1536
    def aggregation(self) -> dict:
1537
1538
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1539
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1540
        return self._higher_is_better
1541

Baber Abbasi's avatar
Baber Abbasi committed
1542
1543
1544
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1545
1546
1547
1548
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1549
1550
1551
1552
1553
1554
1555
1556
    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)},"
            f"num_samples={len(self.eval_docs)})"
        )

1557
1558

class MultipleChoiceTask(Task):
1559
    OUTPUT_TYPE = "loglikelihood"
1560

baberabb's avatar
baberabb committed
1561
    def doc_to_target(self, doc: dict) -> str:
1562
1563
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1564
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1565
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1566
1567
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1568
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1569
                doc=doc,
1570
                arguments=(ctx, " {}".format(choice)),
1571
                idx=i,
1572
1573
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1574
1575
            for i, choice in enumerate(doc["choices"])
        ]
1576

1577
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1578
1579
1580
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
        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
1592
    def higher_is_better(self) -> dict:
1593
1594
1595
1596
1597
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1598
    def aggregation(self) -> dict:
1599
1600
1601
1602
1603
1604
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1605
class PerplexityTask(Task):
1606
1607
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1608
    def has_training_docs(self) -> bool:
1609
1610
        return False

baberabb's avatar
baberabb committed
1611
    def fewshot_examples(self, k: int, rnd) -> List:
1612
1613
1614
1615
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1616
1617
        return []

baberabb's avatar
baberabb committed
1618
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1619
1620
1621
1622
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1623
1624
1625

        return ""

baberabb's avatar
baberabb committed
1626
    def higher_is_better(self) -> dict:
1627
1628
1629
1630
1631
1632
1633
1634
1635
        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
1636
    def doc_to_text(self, doc) -> str:
1637
1638
1639
1640
1641
        return ""

    def doc_to_target(self, doc):
        return doc

1642
1643
1644
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1645

lintangsutawika's avatar
lintangsutawika committed
1646
1647
1648
1649
1650
1651
1652
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1653

1654
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1655
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1656
1657
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1658
1659
1660
1661
1662
1663
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1664
    def aggregation(self) -> dict:
1665
1666
1667
1668
1669
1670
1671
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1672
    def count_bytes(cls, doc) -> int:
1673
1674
1675
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1676
    def count_words(cls, doc) -> int:
1677
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1678
        return len(re.split(r"\s+", doc))