task.py 60.9 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 typing import (
Baber's avatar
Baber committed
9
    TYPE_CHECKING,
10
11
12
13
14
15
16
17
18
19
20
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
21
22
23

import datasets
import numpy as np
24
from tqdm import tqdm
Baber's avatar
Baber committed
25
from typing_extensions import deprecated
26
27

from lm_eval import utils
28
29
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
30
from lm_eval.caching.cache import load_from_cache, save_to_cache
Baber's avatar
Baber committed
31
32
from lm_eval.config.metric import MetricConfig
from lm_eval.config.task import TaskConfig
33
34
35
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

36

37
38
39
40
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
41
    "generate_until",
42
43
]

Baber's avatar
Baber committed
44
if TYPE_CHECKING:
Baber's avatar
Baber committed
45
    pass
Baber's avatar
Baber committed
46
47


Lintang Sutawika's avatar
Lintang Sutawika committed
48
eval_logger = logging.getLogger(__name__)
49

lintangsutawika's avatar
lintangsutawika committed
50

51
52
53
54
55
56
57
58
59
60
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)
    """

61
    VERSION: Optional[Union[int, str]] = None
62

63
64
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
65
    DATASET_PATH: Optional[str] = None
66
67

    # The name of a subset within `DATASET_PATH`.
68
    DATASET_NAME: Optional[str] = None
69

70
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
71

72
73
    def __init__(
        self,
74
75
76
77
        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
78
    ) -> None:
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
        """
        :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)
101
102
103
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
104

105
        self._config: TaskConfig = TaskConfig.from_yaml({**config})
106

107
        self._filters = [build_filter_ensemble("none", [("take_first", None)])]
108
109
110
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
111

112
113
114
115
116
117
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
        """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.
        """
142
143
144
145
146
147
148
        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,
        )
149

150
    @property
151
    def config(self) -> TaskConfig:
152
153
154
        """Returns the TaskConfig associated with this class."""
        return self._config

155
    @abc.abstractmethod
Baber's avatar
Baber committed
156
    def has_training_docs(self) -> bool:
157
158
159
160
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
161
    def has_validation_docs(self) -> bool:
162
163
164
165
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
166
    def has_test_docs(self) -> bool:
167
168
169
        """Whether the task has a test set"""
        pass

170
    def training_docs(self) -> Iterable:
171
172
173
174
175
176
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

177
    def validation_docs(self) -> Iterable:
178
179
180
181
182
183
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

184
    def test_docs(self) -> Iterable:
185
186
187
188
189
190
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

191
    def fewshot_docs(self) -> Iterable:
192
193
194
195
196
197
198
199
200
        """
        :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:
Baber Abbasi's avatar
Baber Abbasi committed
201
202
203
204
205
            if self.config.get("num_fewshot", 0) > 0:
                eval_logger.warning(
                    f"[Task: {self.config.task}] has_training_docs and has_validation_docs are False"
                    ", using test_docs as fewshot_docs but this is not recommended."
                )
206
207
            return self.test_docs()

208
    def _process_doc(self, doc: dict) -> dict:
209
210
211
212
213
214
215
216
217
        """
        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
218

219
    @property
Baber's avatar
Baber committed
220
    def instances(self) -> list[Instance]:
221
222
223
224
225
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

Baber's avatar
Baber committed
226
    def fewshot_examples(self, k, rnd) -> Iterable[dict]:
227
228
229
230
231
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

Baber's avatar
Baber committed
232
    def doc_to_decontamination_query(self, doc: dict):
233
        raise NotImplementedError(
234
235
236
237
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
Baber committed
238
    def doc_to_text(self, doc: dict) -> str:
239
240
241
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
242
    def doc_to_target(self, doc: dict) -> Union[str, int]:
243
244
        pass

245
    # not an abstractmethod because not every language-only task has to implement this
Baber's avatar
Baber committed
246
    def doc_to_image(self, doc: dict):
247
248
        raise NotImplementedError

Baber's avatar
Baber committed
249
    def doc_to_audio(self, doc: dict):
250
251
        raise NotImplementedError

Baber's avatar
Baber committed
252
    def doc_to_prefix(self, doc: dict) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
253
254
        return ""

255
256
    def build_all_requests(
        self,
257
        *,
258
        limit: Union[int, None] = None,
259
        samples: Optional[List[int]] = None,
260
261
262
263
264
265
266
267
268
        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 = "",
269
    ) -> None:
270
        """Build a set of Instances for a task, and store them in task.instances"""
271
272
273
274

        # used with caching
        og_limit = limit

275
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
276
277
278
279
280
281
282
        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 ""
        )
283
        cache_key += f"-tokenizer{tokenizer_name}"
284

Baber Abbasi's avatar
Baber Abbasi committed
285
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
286
287
288
289
290
291
292
293
294
295
296
297
298

        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
299
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
300

301
        instances = []
302
303
304
305
306
307
308
309
310
311

        # 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(
312
313
314
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
315
316
317
318
319
320
321
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
322
        ):
323
            # sample fewshot context #TODO: need to offset doc_id by rank now!
324
            fewshot_ctx = self.fewshot_context(
325
                doc,
326
327
328
329
330
331
332
                num_fewshot=0
                if self.config.num_fewshot is None
                else self.config.num_fewshot,
                system_instruction=system_instruction,
                apply_chat_template=apply_chat_template,
                fewshot_as_multiturn=fewshot_as_multiturn,
                chat_template=chat_template,
Baber Abbasi's avatar
Baber Abbasi committed
333
                gen_prefix=self.doc_to_prefix(doc),
334
            )
335

336
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
337
338
339
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
340
                metadata=(self.config["task"], doc_id, self.config.repeats),
341
                apply_chat_template=apply_chat_template,
342
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
343
            )
344
345
346
347

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

348
349
350
351
352
353
354
355
356
357
358
359
360
            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
361

362
363
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
364

365
366
367
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

368
    @abc.abstractmethod
Baber's avatar
Baber committed
369
    def construct_requests(self, doc: dict, ctx: Union[list[dict], str], **kwargs):
370
371
372
373
374
375
376
377
378
379
380
381
382
383
        """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
384
            The number of times each instance in a dataset is inferred on. Defaults to 1,
385
386
387
388
389
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
390
    def process_results(self, doc: dict, results: list):
391
392
393
394
395
396
397
398
399
400
401
        """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

Baber's avatar
Baber committed
402
    @deprecated("not used anymore")
403
404
405
406
407
408
409
410
    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

Baber's avatar
Baber committed
411
    @deprecated("not used anymore")
412
413
414
415
416
417
418
419
    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

420
421
422
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
423
    @classmethod
Baber's avatar
Baber committed
424
    def count_bytes(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
425
426
427
428
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
429
    def count_words(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
430
431
432
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

433
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
434
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
435
436
437
438
439
440
441
        """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
442
443
444
445
446
        :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.
447
448
449
        :returns: str
            The fewshot context.
        """
450
        if rnd is None:
451
452
453
454
455
456
            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
457

458
        description = description if description else ""
459
460

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
461
            labeled_examples = ""
462
        else:
lintangsutawika's avatar
lintangsutawika committed
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
            # 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
487
            )
488
489

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
490
        return description + labeled_examples + example
491

492
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
493
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
494
495
        if hasattr(self, "_filters"):
            for f in self._filters:
496
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
497
498
499
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
500

baberabb's avatar
baberabb committed
501
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
502
        """Returns the config as a dictionary."""
503
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
504
        # (num_fewshot)
505
        return self.config.to_dict()
506

Baber Abbasi's avatar
Baber Abbasi committed
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
    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.
        """
Baber's avatar
Baber committed
529
530
531
532
533
534
535
        # 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", [MetricConfig(name=metric_name)])
        setattr(self._config, "process_results", lambda *args: {"bypass": 0})
Baber Abbasi's avatar
Baber Abbasi committed
536

537
538
539
540
541
    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

542
    @property
Baber's avatar
Baber committed
543
    def eval_docs(self) -> Union[datasets.Dataset, Iterable[dict]]:
544
545
546
547
548
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
549
550
551
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
552
553

    def doc_iterator(
554
555
556
557
558
559
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
560
    ) -> Iterator[Tuple[int, Any]]:
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
        if samples:
            n = len(self.eval_docs)
            assert all([e < n for e in samples]), (
                f"Elements of --samples should be in the interval [0,k-1] where k is the number of total examples. In this case, k={n}."
            )
            eval_logger.info(
                f"{self.config.task}: Evaluating on {len(samples)} examples"
            )
            doc_iterator = utils.create_iterator(
                enumerate(x for i, x in enumerate(self.eval_docs) if i in samples),
                rank=int(rank),
                limit=None,  # limit does not matter here since we are selecting samples directly
                world_size=int(world_size),
            )
        else:
            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),
            )
583
584
        return doc_iterator

585
586

class ConfigurableTask(Task):
587
    VERSION = "Yaml"
588
    OUTPUT_TYPE = None
589
    CONFIG = None
590
591

    def __init__(
592
593
594
595
596
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Baber's avatar
Baber committed
597
    ) -> None:
598
        # Get pre-configured attributes
599
        self._config = self.CONFIG
600

601
        # Use new configurations if there was no preconfiguration
602
        if self.config is None:
603
            self._config = TaskConfig(**config)
604
605
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
606
            if config is not None:
607
                self._config.__dict__.update(config)
608

609
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
610
611
612
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
613

614
615
616
617
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

618
        if self.config.output_type is not None:
619
620
621
622
            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)}'"
                )
623
            self.OUTPUT_TYPE = self.config.output_type
624

625
626
627
628
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

629
630
631
632
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
633
634
635
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

636
637
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
638

639
640
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
641

642
        # self.metric_list: list[MetricConfig] = self.config.get_metrics
643

644
        self.download(self.config.dataset_kwargs)
645
646
647
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
648
        self._filters = self.config.get_filters
Baber's avatar
Baber committed
649

650
651
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
652
            self.prompt = get_prompt(
653
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
654
            )
655
656
657
        else:
            self.prompt = None

658
659
660
661
        if (
            self.config.fewshot_cfg.num_fewshot() > 0
            and self.fewshot_docs() is not None
        ):
Baber's avatar
Baber committed
662
663
664
            self.fewshot_rnd = random.Random()
            self.sampler = self.config.fewshot_cfg.init_sampler(
                list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
665
            )
666
        self.task_docs = self.eval_docs
667

668
        # Test One Doc
669
        self.features = list(self.task_docs.features.keys())
670
671
        self.multiple_input = 0
        self.multiple_target = 0
672
        test_doc = self.task_docs[0]
673
        test_text = self.doc_to_text(test_doc)
674
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
675

676
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
677
            test_choice = self.doc_to_choice(test_doc)
678
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
679
                eval_logger.error("doc_to_choice must return list")
680
681
            else:
                num_choice = len(test_choice)
682

683
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
684
685
686
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
687
                self.multiple_input = num_choice
688
689
        else:
            test_choice = None
690

691
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
692
693
694
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
695
            self.multiple_target = len(test_target)
696
        else:
697
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
698
                test_target = test_choice[test_target]
699
            else:
lintangsutawika's avatar
lintangsutawika committed
700
                test_target = str(test_target)
701

702
703
704
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
705
            check_choices = [test_target]
706
707
708
709
        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 = (
710
711
                    True
                    if self.config.target_delimiter.rstrip()
712
                    != self.config.target_delimiter
713
                    else False
714
                )
715

716
                if delimiter_has_whitespace and choice_has_whitespace:
717
718
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
719
720
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
721
                    eval_logger.debug(
722
                        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'
723
724
                    )

Baber Abbasi's avatar
Baber Abbasi committed
725
726
727
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
728
729
730
731
732
        self.config.dataset_kwargs, self.config.metadata = (
            self.config.dataset_kwargs or {},
            self.config.metadata or {},
        )
        if isinstance(df := self.config.custom_dataset, Callable):
Baber Abbasi's avatar
Baber Abbasi committed
733
734
735
736
            eval_logger.warning(
                f"{self.config.task}: Custom kwargs can be passed to `--metadata` in console (as json string) or to the TaskManager."
                + "\nFor example --metadata='{\"max_seq_lengths\":[4096, 8192]}'. For details see task Readme."
            )
737
            self.dataset = df(**(self.config.dataset_kwargs | self.config.metadata))
Baber Abbasi's avatar
Baber Abbasi committed
738
739
        else:
            self.dataset = datasets.load_dataset(
740
741
742
                path=self.config.dataset_path,
                name=self.config.dataset_name,
                **self.config.dataset_kwargs,
Baber Abbasi's avatar
Baber Abbasi committed
743
            )
744

baberabb's avatar
baberabb committed
745
    def has_training_docs(self) -> bool:
746
        if self.config.training_split is not None:
747
748
749
750
            return True
        else:
            return False

baberabb's avatar
baberabb committed
751
    def has_validation_docs(self) -> bool:
752
        if self.config.validation_split is not None:
753
754
755
756
            return True
        else:
            return False

baberabb's avatar
baberabb committed
757
    def has_test_docs(self) -> bool:
758
        if self.config.test_split is not None:
759
760
761
762
            return True
        else:
            return False

Baber's avatar
Baber committed
763
    def training_docs(self) -> Optional[datasets.Dataset]:
764
        if self.has_training_docs():
765
766
767
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
768
                )
769
            return self.dataset[self.config.training_split]
770

Baber's avatar
Baber committed
771
    def validation_docs(self) -> Optional[datasets.Dataset]:
772
        if self.has_validation_docs():
773
774
775
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
776
                )
777
            return self.dataset[self.config.validation_split]
778

Baber's avatar
Baber committed
779
    def test_docs(self) -> Optional[datasets.Dataset]:
780
        if self.has_test_docs():
781
782
783
            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]
784

785
    def fewshot_docs(self):
Baber's avatar
Baber committed
786
787
788
789
790
791
792
793
        docs = self.config.fewshot_cfg.get_docs(self.dataset)

        if docs is not None:
            return docs

        # Fallback to parent implementation
        if _num_fewshot := getattr(self.config, "num_fewshot"):
            if isinstance(_num_fewshot, int) and _num_fewshot > 0:
794
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
795
                    f"[Task: {self.config.task}] "
Baber's avatar
Baber committed
796
797
                    "num_fewshot > 0 but no fewshot source configured. "
                    "Using preconfigured rule."
798
                )
Baber's avatar
Baber committed
799
800

        return super().fewshot_docs()
801

KonradSzafer's avatar
KonradSzafer committed
802
803
804
805
806
    @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
807
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
808
809
810
811
812
813
814
815
    ) -> 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":
816
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
817
818
            # if last message is user, append to it to avoid two user messages in a row
            else:
819
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
820
821
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
822
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
823
824
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
825

lintangsutawika's avatar
lintangsutawika committed
826
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
827
828
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
829
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
830
831
832
833
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
834
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
835
        gen_prefix: Optional[str] = None,
Baber's avatar
Baber committed
836
    ) -> Union[str, List[str], None]:
lintangsutawika's avatar
lintangsutawika committed
837
838
839
840
841
842
843
        """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
844
845
846
847
848
849
        :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.
850
851
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
852
853
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
854
855
856
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
857
858
859
860
861
862
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
863
864
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
865

KonradSzafer's avatar
KonradSzafer committed
866
867
868
869
870
871
872
873
874
        # 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
875
        else:
KonradSzafer's avatar
KonradSzafer committed
876
877
878
879
880
881
882
883
884
885
886
887
888
            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
889
890
891
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
892
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
893
894
895
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
896
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
897
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
898
                )
lintangsutawika's avatar
lintangsutawika committed
899
900

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
901
902
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
903
                # TODO: append prefill?
904
905
                if not labeled_examples:
                    return ""
906
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
907
908
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
909
910
911
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
912
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
913
914
915
916
917
918
919
                )
            # 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
920
921
922
923
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
924
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
925
926
927
928
929
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
930
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
931
932
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
933
934
935
936
937
938
                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
939
940
941
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
942
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
943
944
945
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
946
947
948
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
949
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
950
951
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
952
953
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
954
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
955
            )
956
        else:
Baber Abbasi's avatar
Baber Abbasi committed
957
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
958
959
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
960
961
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
962
963
            if self.multiple_input:
                return labeled_examples
964
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
965
                return labeled_examples + example + prefix
966
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
967
                return [labeled_examples + ex + prefix for ex in example]
968
969
970
            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
971
                    return labeled_examples + choices[example] + prefix
972
                else:
Baber Abbasi's avatar
Baber Abbasi committed
973
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
974

Baber Abbasi's avatar
Baber Abbasi committed
975
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
976
        """Iterates over FilterEnsembles and applies them to instances"""
977
978
        if hasattr(self, "_filters"):
            for f in self._filters:
979
                f.ensemble.apply(self._instances)
980
981
982
983
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

984
    def should_decontaminate(self):
985
        return self.config.should_decontaminate
986

Baber Abbasi's avatar
Baber Abbasi committed
987
    def doc_to_decontamination_query(self, doc: dict):
988
        if self.config.should_decontaminate:
989
990
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
991
            else:
992
993
994
995
996
997
998
999
1000
1001
1002
                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
                        )
                    )
1003

1004
    def _process_doc(self, doc: dict) -> dict:
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
        """
        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

Baber's avatar
Baber committed
1015
    def doc_to_text(self, doc: dict, doc_to_text: Union[int, str, Callable] = None):
1016
1017
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1018
1019
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1020
        else:
1021
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1022

1023
        if isinstance(doc_to_text, int):
1024
            return doc_to_text
1025
        elif isinstance(doc_to_text, str):
1026
            if doc_to_text in self.features:
1027
                # if self.config.doc_to_choice is not None:
1028
1029
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1030
1031
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1032
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
nit  
Baber committed
1033
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1034
1035
1036
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1037
        elif callable(doc_to_text):
1038
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1039
        # Used when applying a Promptsource template
1040
        elif hasattr(doc_to_text, "apply"):
1041
1042
1043
1044
1045
            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")
1046
                return self.config.fewshot_delimiter
1047
        else:
1048
            print(type(doc_to_text))
1049
            raise TypeError
1050

Baber's avatar
Baber committed
1051
    def doc_to_target(self, doc: dict, doc_to_target=None) -> Union[int, str, list]:
1052
1053
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1054
1055
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1056
        else:
1057
            doc_to_target = self.config.doc_to_target
1058

1059
        if isinstance(doc_to_target, int):
1060
            return doc_to_target
1061
        elif isinstance(doc_to_target, str):
1062
            if doc_to_target in self.features:
1063
                # if self.config.doc_to_choice is not None:
1064
1065
1066
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1067
            else:
lintangsutawika's avatar
lintangsutawika committed
1068
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
nit  
Baber committed
1069
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1070
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1071
1072
1073
1074
1075
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1076
1077
1078
1079
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1080
1081
                else:
                    return target_string
1082
        elif isinstance(doc_to_target, list):
1083
            return doc_to_target
1084
        elif callable(doc_to_target):
1085
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1086
        # Used when applying a Promptsource template
1087
        elif hasattr(doc_to_target, "apply"):
1088
            applied_prompt = doc_to_target.apply(doc)
1089
1090
1091
1092
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1093
                return self.config.fewshot_delimiter
1094
1095
        else:
            raise TypeError
1096

Baber's avatar
Baber committed
1097
1098
1099
    def doc_to_choice(
        self, doc: dict, doc_to_choice: Union[str, list, dict] = None
    ) -> List[str]:
1100
1101
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1102
1103
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1104
        elif self.config.doc_to_choice is None:
1105
1106
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1107
            doc_to_choice = self.config.doc_to_choice
1108

1109
        if isinstance(doc_to_choice, str):
1110
1111
1112
1113
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1114
        elif isinstance(doc_to_choice, list):
1115
            return doc_to_choice
1116
        elif isinstance(doc_to_choice, dict):
1117
1118
1119
1120
1121
1122
1123
            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
1124

Baber's avatar
Baber committed
1125
    def doc_to_image(self, doc: dict, doc_to_image=None) -> Union[int, str, list, None]:
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
        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's avatar
Baber committed
1148
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list, None]:
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber's avatar
Baber committed
1171
    def doc_to_prefix(self, doc: dict) -> Optional[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1172
1173
1174
1175
1176
1177
1178
        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
1179
1180
1181
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1182
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1183
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1184

1185
1186
        aux_arguments = None

1187
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1188
            arguments = (ctx, self.doc_to_target(doc))
1189
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1190
            arguments = (self.doc_to_target(doc),)
1191
        elif self.OUTPUT_TYPE == "multiple_choice":
1192
            choices = self.doc_to_choice(doc)
1193
            target_delimiter = self.config.target_delimiter
1194
1195
            if apply_chat_template:
                target_delimiter = ""
1196
1197
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1198
                # apply chat_template to choices if apply_chat_template
1199
                cont = self.doc_to_target(doc)
1200

1201
                arguments = [
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1212
                ]
1213
            else:
1214
                # Otherwise they are placed in the continuation
1215
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1216

1217
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
Baber's avatar
Baber committed
1218
            if "acc_mutual_info" in [m.metric_name for m in self.metric_list]:
1219
1220
1221
1222
1223
1224
                # 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.
1225
1226
1227
1228
                # TODO: should these be strided? will have to modify the processing in process_results if so
                aux_arguments = [
                    ("", f"{target_delimiter}{choice}") for choice in choices
                ]
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243

                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)},
            }

1244
1245
1246
1247
1248
1249
1250
1251
        if (
            self.config.doc_to_audio
        ):  # TODO: ensure that non-multimodal tasks aren't getting audio args
            multimodal_arg = {
                **multimodal_arg,
                **{"audio": self.doc_to_audio(doc)},
            }

1252
1253
1254
1255
1256
1257
1258
        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":
1259
            request_list = [
1260
1261
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1262
                    doc=doc,
1263
                    arguments=arg,
1264
                    idx=i,
1265
1266
                    **kwargs,
                )
1267
                for i, arg in enumerate(arguments)
1268
            ]
1269
1270

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1271

lintangsutawika's avatar
lintangsutawika committed
1272
        return Instance(
1273
1274
1275
1276
1277
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1278
        )
1279

Baber's avatar
Baber committed
1280
    def process_results(self, doc: dict, results: list) -> dict:
1281
1282
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1283

1284
        result_dict = {}
Baber's avatar
fixup  
Baber committed
1285
        use_metric = list(m.metric_name for m in self.config._metric_list)
1286
1287
1288
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1289
1290
1291
1292
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1293
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1294
            (loglikelihood,) = results
1295
1296
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1297
            return {
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
                **(
                    {"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
1313
            }
1314
        elif self.OUTPUT_TYPE == "multiple_choice":
1315
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1316

1317
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1318
            choices = self.doc_to_choice(doc)
1319
1320
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1321
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1322
1323
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1324
1325
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1326
1327
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1328
                # and this stores our "regular" conditional loglikelihoods
1329
                lls = lls[: len(choices)]
1330

1331
1332
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1333

1334
1335
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1336
            else:
1337
                gold = self.doc_to_target(doc)
1338
1339

            gold_index_error = False
1340
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1341
1342
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1343
1344
                    gold_index_error = True
            else:
1345
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1346
                    gold = gold if gold < len(choices) else -100
1347
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1348
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1349

Lintang Sutawika's avatar
Lintang Sutawika committed
1350
                if gold == -100:
1351
1352
1353
1354
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1355
                    f"Label index was not in within range of available choices,"
1356
1357
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1358

1359
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1360
1361
                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
1362
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1363
1364
1365
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1366
                # 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
1367
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1368

Lintang Sutawika's avatar
Lintang Sutawika committed
1369
1370
1371
1372
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1373
            result_dict = {
1374
                **({"acc": acc} if "acc" in use_metric else {}),
1375
1376
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1377
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1378
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1379
1380
1381
1382
1383
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1384
1385
            }

1386
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1387
1388
1389
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1390
1391
1392
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1393
        elif self.OUTPUT_TYPE == "generate_until":
1394
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1395
            result = results[0]
1396
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1397
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1398
                # it assumes that doc_to_target returns a number.
1399
1400
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1401
1402
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1403
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1404
1405
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
Baber's avatar
Baber committed
1406
                "bypass" in use_metric or isinstance(result, list)
1407
            ):
Chris's avatar
Chris committed
1408
1409
                # cast gold to the same type as result
                gold = type(result)(gold)
1410

Baber's avatar
fixup  
Baber committed
1411
            for metric in self.config._metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
1412
1413
1414
1415
1416
                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
1417
1418
1419
1420
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
Baber's avatar
Baber committed
1421
                    if metric.name == "exact_match":
1422
                        result = [result for _ in range(len(gold))]
Baber's avatar
Baber committed
1423
                        scores = metric.fn(
1424
1425
                            references=gold,
                            predictions=result,
Baber's avatar
Baber committed
1426
                            **metric.kwargs,
1427
1428
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1429
                    else:
1430
1431
                        for gold_option in gold:
                            try:
Baber's avatar
Baber committed
1432
                                result_score = metric.fn(
1433
1434
                                    references=[gold_option],
                                    predictions=[result],
Baber's avatar
Baber committed
1435
                                    **metric.kwargs,
1436
1437
1438
1439
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
Baber's avatar
Baber committed
1440
                                result_score = metric.fn([gold_option, result])
1441
1442
1443
1444
1445
1446
1447
1448
                            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
1449
                else:
1450
                    try:
Baber's avatar
Baber committed
1451
                        result_score = metric.fn(
1452
1453
                            references=[gold],
                            predictions=[result],
Baber's avatar
Baber committed
1454
                            **metric.kwargs,
1455
                        )
1456
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
Baber's avatar
Baber committed
1457
                        result_score = metric.fn([gold, result])
1458
1459
1460
1461
1462
1463
1464
                if isinstance(result_score, dict):
                    # TODO: this handles the case where HF evaluate returns a dict.
                    # This allows for multiple metrics to be returned from the same function
                    for k, v in result_score.items():
                        result_dict[k] = v
                else:
                    result_dict[metric] = result_score
1465
        else:
lintangsutawika's avatar
lintangsutawika committed
1466
1467
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1468
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1469
            )
1470
1471
1472

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1473
    def aggregation(self) -> dict:
Baber's avatar
fixup  
Baber committed
1474
        return {k.name: k.aggregation_fn for k in self.config._metric_list}
1475

Baber Abbasi's avatar
Baber Abbasi committed
1476
    def higher_is_better(self) -> dict:
Baber's avatar
fixup  
Baber committed
1477
        return {k.name: k.higher_is_better for k in self.config._metric_list}
1478

Baber Abbasi's avatar
Baber Abbasi committed
1479
1480
1481
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1482
    @property
Baber's avatar
Baber committed
1483
    def task_name(self) -> Optional[str]:
Lintang Sutawika's avatar
Lintang Sutawika committed
1484
1485
        return getattr(self.config, "task", None)

1486
1487
1488
1489
1490
    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
1491
            f"num_samples={len(self.eval_docs)})"
1492
1493
        )

1494
1495

class MultipleChoiceTask(Task):
1496
    OUTPUT_TYPE = "loglikelihood"
1497

baberabb's avatar
baberabb committed
1498
    def doc_to_target(self, doc: dict) -> str:
1499
1500
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1501
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1502
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1503
1504
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1505
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1506
                doc=doc,
1507
                arguments=(ctx, " {}".format(choice)),
1508
                idx=i,
1509
1510
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1511
1512
            for i, choice in enumerate(doc["choices"])
        ]
1513

1514
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1515
1516
1517
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
        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
1529
    def higher_is_better(self) -> dict:
1530
1531
1532
1533
1534
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1535
    def aggregation(self) -> dict:
1536
1537
1538
1539
1540
1541
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1542
class PerplexityTask(Task):
1543
1544
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1545
    def has_training_docs(self) -> bool:
1546
1547
        return False

baberabb's avatar
baberabb committed
1548
    def fewshot_examples(self, k: int, rnd) -> List:
1549
1550
1551
1552
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1553
1554
        return []

baberabb's avatar
baberabb committed
1555
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1556
1557
1558
1559
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1560
1561
1562

        return ""

baberabb's avatar
baberabb committed
1563
    def higher_is_better(self) -> dict:
1564
1565
1566
1567
1568
1569
1570
1571
1572
        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
1573
    def doc_to_text(self, doc) -> str:
1574
1575
1576
1577
1578
        return ""

    def doc_to_target(self, doc):
        return doc

1579
1580
1581
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1582

lintangsutawika's avatar
lintangsutawika committed
1583
1584
1585
1586
1587
1588
1589
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1590

1591
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1592
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1593
1594
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1595
1596
1597
1598
1599
1600
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1601
    def aggregation(self) -> dict:
1602
1603
1604
1605
1606
1607
1608
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1609
    def count_bytes(cls, doc) -> int:
1610
1611
1612
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1613
    def count_words(cls, doc) -> int:
1614
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1615
        return len(re.split(r"\s+", doc))