task.py 72.6 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
import re
Baber's avatar
Baber committed
6
from collections.abc import Callable, Iterable, Iterator, Mapping
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
from typing import (
    Any,
    Dict,
    List,
    Literal,
    Optional,
    Tuple,
    Union,
)
19
20
21

import datasets
import numpy as np
22
from tqdm import tqdm
23
24

from lm_eval import utils
25
from lm_eval.api import samplers
26
27
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
28
from lm_eval.api.registry import (
29
30
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
31
    get_aggregation,
32
    get_metric,
33
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
34
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
35
)
36
from lm_eval.caching.cache import load_from_cache, save_to_cache
37
38
39
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

40

41
42
43
44
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
45
    "generate_until",
46
47
]

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

lintangsutawika's avatar
lintangsutawika committed
50

51
52
@dataclass
class TaskConfig(dict):
53
    # task naming/registry
54
55
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
56
    tag: Optional[Union[str, list]] = None
57
58
59
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
60
    custom_dataset: Optional[Callable] = None
61
62
63
64
65
66
    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
67
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
68
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
69
    )
70
71
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
72
73
74
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
75
    doc_to_image: Union[Callable, str] = None
76
    doc_to_audio: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
77
    unsafe_code: bool = False
78
79
80
    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
Baber Abbasi's avatar
Baber Abbasi committed
95
    gen_prefix: Optional[str] = None
96
97
98
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
99

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

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

            if "until" not in self.generation_kwargs:
Baber Abbasi's avatar
Baber Abbasi committed
113
114
115
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
116
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
117
        else:
118
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
119
120
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
121
122
123
124
125
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
126
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
127
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
128
                }
Baber Abbasi's avatar
Baber Abbasi committed
129
130
131
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
132

133
134
135
    def __getitem__(self, item):
        return getattr(self, item)

136
137
138
    def __setitem__(self, item, value):
        return setattr(self, item, value)

139
    def to_dict(self, keep_callable: bool = False) -> dict:
140
141
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
142
        Used for dumping results alongside full task configuration
143

haileyschoelkopf's avatar
haileyschoelkopf committed
144
145
146
147
148
149
150
151
152
153
        :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)
154
155
156
157
158
159
160
161
162
163
            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
164
        return cfg_dict
165

166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
    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)

182
183
184
185
186
187
188
189
190
191
192

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)
    """

193
    VERSION: Optional[Union[int, str]] = None
194

195
196
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
197
    DATASET_PATH: Optional[str] = None
198
199

    # The name of a subset within `DATASET_PATH`.
200
    DATASET_NAME: Optional[str] = None
201

202
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
203

204
205
    def __init__(
        self,
206
207
208
209
        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
210
    ) -> None:
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
        """
        :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)
233
234
235
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
236

237
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
238

lintangsutawika's avatar
lintangsutawika committed
239
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
240
241
242
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
243

244
245
246
247
248
249
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
        """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.
        """
274
275
276
277
278
279
280
        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,
        )
281

282
    @property
283
    def config(self) -> TaskConfig:
284
285
286
        """Returns the TaskConfig associated with this class."""
        return self._config

287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
    @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

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

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

316
    def test_docs(self) -> Iterable:
317
318
319
320
321
322
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

323
    def fewshot_docs(self) -> Iterable:
324
325
326
327
328
329
330
331
332
        """
        :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
333
334
335
336
337
            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."
                )
338
339
            return self.test_docs()

340
    def _process_doc(self, doc: dict) -> dict:
341
342
343
344
345
346
347
348
349
        """
        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
350

351
    @property
352
    def instances(self) -> List[Instance]:
353
354
355
356
357
358
359
360
361
362
363
        """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)

364
365
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
366
367
368
369
370
371
372
373
374
375
376
            "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

377
378
379
380
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

381
382
383
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
384
385
386
    def doc_to_prefix(self, doc):
        return ""

387
388
    def build_all_requests(
        self,
389
        *,
390
        limit: Union[int, None] = None,
391
        samples: Optional[List[int]] = None,
392
393
394
395
396
397
398
399
400
        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 = "",
401
    ) -> None:
402
        """Build a set of Instances for a task, and store them in task.instances"""
403
404
405
406

        # used with caching
        og_limit = limit

407
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
408
409
410
411
412
413
414
        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 ""
        )
415
        cache_key += f"-tokenizer{tokenizer_name}"
416

Baber Abbasi's avatar
Baber Abbasi committed
417
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
418
419
420
421
422
423
424
425
426
427
428
429
430

        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
431
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
432

433
        instances = []
434
435
436
437
438
439
440
441
442
443

        # 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(
444
445
446
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
447
448
449
450
451
452
453
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
454
        ):
455
            # sample fewshot context #TODO: need to offset doc_id by rank now!
456
            fewshot_ctx = self.fewshot_context(
457
                doc,
458
459
460
461
462
463
464
                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
465
                gen_prefix=self.doc_to_prefix(doc),
466
            )
467

468
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
469
470
471
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
472
                metadata=(self.config["task"], doc_id, self.config.repeats),
473
                apply_chat_template=apply_chat_template,
474
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
475
            )
476
477
478
479

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

480
481
482
483
484
485
486
487
488
489
490
491
492
            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
493

494
495
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
496

497
498
499
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
    @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
516
            The number of times each instance in a dataset is inferred on. Defaults to 1,
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
            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

552
553
554
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
555
556
557
558
559
560
561
562
563
564
    @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))

565
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
566
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
567
568
569
570
571
572
573
        """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
574
575
576
577
578
        :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.
579
580
581
        :returns: str
            The fewshot context.
        """
582
        if rnd is None:
583
584
585
586
587
588
            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
589

590
        description = description if description else ""
591
592

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
593
            labeled_examples = ""
594
        else:
lintangsutawika's avatar
lintangsutawika committed
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
            # 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
619
            )
620
621

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
622
        return description + labeled_examples + example
623

624
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
625
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
626
627
        if hasattr(self, "_filters"):
            for f in self._filters:
628
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
629
630
631
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
632

baberabb's avatar
baberabb committed
633
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
634
        """Returns the config as a dictionary."""
635
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
636
        # (num_fewshot)
637
        return self.config.to_dict()
638

Baber Abbasi's avatar
Baber Abbasi committed
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
671
672
673
674
675
676
677
678
    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)

679
680
681
682
683
    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

684
685
686
687
688
689
690
    @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:
691
692
693
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
694
695

    def doc_iterator(
696
697
698
699
700
701
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
702
    ) -> Iterator[Tuple[int, Any]]:
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
        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),
            )
725
726
        return doc_iterator

727
728

class ConfigurableTask(Task):
729
    VERSION = "Yaml"
730
    OUTPUT_TYPE = None
731
    CONFIG = None
732
733

    def __init__(
734
735
736
737
738
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
739
    ) -> None:  # TODO no super() call here
740
        # Get pre-configured attributes
741
        self._config = self.CONFIG
742

743
        # Use new configurations if there was no preconfiguration
744
        if self.config is None:
745
            self._config = TaskConfig(**config)
746
747
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
748
            if config is not None:
749
                self._config.__dict__.update(config)
750

751
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
752
753
754
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
755

756
757
758
759
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

760
        if self.config.output_type is not None:
761
762
763
764
            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)}'"
                )
765
            self.OUTPUT_TYPE = self.config.output_type
766

767
768
769
770
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

771
772
773
774
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
775
776
777
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

778
779
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
780

781
782
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
783

784
785
786
787
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
788

789
        if self.config.metric_list is None:
790
            # TODO: handle this in TaskConfig.__post_init__ ?
791
792
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

793
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
794
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
795
                self._metric_fn_kwargs[metric_name] = {}
796
797
798
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
799
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
800
        else:
801
            for metric_config in self.config.metric_list:
802
803
804
805
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
806
807
808
809
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
810
811
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
812
                }
Chris's avatar
Chris committed
813
814
815
816
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
817

818
                if self.config.process_results is not None:
819
820
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
821
822
823
824
825
826
                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
827
828
829
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
830
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
831

832
                if "aggregation" in metric_config:
833
                    agg_name = metric_config["aggregation"]
834
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
835
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
836
                    elif callable(agg_name):  # noqa: E721
837
838
839
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
840
                else:
841
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
842
                    metric_agg = get_metric_aggregation(metric_name)
843
                    eval_logger.warning(
844
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
845
846
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
847
                    )
848
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
849

850
851
852
853
854
855
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
856
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
857
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
858
                        f"higher_is_better={is_higher_better(metric_name)}"
859
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
860
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
861

862
        self.download(self.config.dataset_kwargs)
863
864
865
        self._training_docs = None
        self._fewshot_docs = None

866
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
867
            self._filters = []
868
            for filter_config in self.config.filter_list:
869
870
871
872
873
874
875
876
877
                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
878
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
879
        else:
Baber Abbasi's avatar
Baber Abbasi committed
880
881
882
883
            # TODO: handle repeats in a more general way rather than just discarding
            eval_logger.debug(
                "No custom filters defined. Using default 'take_first' filter for handling repeats."
            )
884
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
885

886
887
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
888
            self.prompt = get_prompt(
889
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
890
            )
891
892
893
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
894
        if self.fewshot_docs() is not None:
895
896
897
898
            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
899
900
901
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
            )
            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)}"
                )
918

919
        self.task_docs = self.eval_docs
920

921
        # Test One Doc
922
        self.features = list(self.task_docs.features.keys())
923
924
        self.multiple_input = 0
        self.multiple_target = 0
925
        test_doc = self.task_docs[0]
926
        test_text = self.doc_to_text(test_doc)
927
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
928

929
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
930
            test_choice = self.doc_to_choice(test_doc)
931
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
932
                eval_logger.error("doc_to_choice must return list")
933
934
            else:
                num_choice = len(test_choice)
935

936
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
937
938
939
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
940
                self.multiple_input = num_choice
941
942
        else:
            test_choice = None
943

944
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
945
946
947
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
948
            self.multiple_target = len(test_target)
949
        else:
950
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
951
                test_target = test_choice[test_target]
952
            else:
lintangsutawika's avatar
lintangsutawika committed
953
                test_target = str(test_target)
954

955
956
957
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
958
            check_choices = [test_target]
959
960
961
962
        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 = (
963
964
                    True
                    if self.config.target_delimiter.rstrip()
965
                    != self.config.target_delimiter
966
                    else False
967
                )
968

969
                if delimiter_has_whitespace and choice_has_whitespace:
970
971
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
972
973
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
974
                    eval_logger.debug(
975
                        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'
976
977
                    )

Baber Abbasi's avatar
Baber Abbasi committed
978
979
980
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
Baber Abbasi's avatar
Baber Abbasi committed
981
982
983
984
        from packaging.version import parse as vparse

        if dataset_kwargs and vparse(datasets.__version__) >= vparse("4.0.0"):
            dataset_kwargs.pop("trust_remote_code", None)
Baber Abbasi's avatar
Baber Abbasi committed
985
986
987
988
989
990
991
992
993
994
995
996
997
998
        if isinstance(self.config.custom_dataset, Callable):
            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."
            )
            self.dataset = self.config.custom_dataset(
                **(self.config.metadata or {}), **(self.config.dataset_kwargs or {})
            )
        else:
            self.dataset = datasets.load_dataset(
                path=self.DATASET_PATH,
                name=self.DATASET_NAME,
                **dataset_kwargs if dataset_kwargs is not None else {},
            )
999

baberabb's avatar
baberabb committed
1000
    def has_training_docs(self) -> bool:
1001
        if self.config.training_split is not None:
1002
1003
1004
1005
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1006
    def has_validation_docs(self) -> bool:
1007
        if self.config.validation_split is not None:
1008
1009
1010
1011
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1012
    def has_test_docs(self) -> bool:
1013
        if self.config.test_split is not None:
1014
1015
1016
1017
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1018
    def training_docs(self) -> datasets.Dataset:
1019
        if self.has_training_docs():
1020
1021
1022
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1023
                )
1024
            return self.dataset[self.config.training_split]
1025

baberabb's avatar
baberabb committed
1026
    def validation_docs(self) -> datasets.Dataset:
1027
        if self.has_validation_docs():
1028
1029
1030
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1031
                )
1032
            return self.dataset[self.config.validation_split]
1033

baberabb's avatar
baberabb committed
1034
    def test_docs(self) -> datasets.Dataset:
1035
        if self.has_test_docs():
1036
1037
1038
            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]
1039

1040
    def fewshot_docs(self):
1041
        if self.config.fewshot_split is not None:
1042
1043
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1044
            return self.dataset[self.config.fewshot_split]
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
        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."
                )
1057
        else:
1058
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1059
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1060
                    f"[Task: {self.config.task}] "
1061
1062
1063
1064
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1065

KonradSzafer's avatar
KonradSzafer committed
1066
1067
1068
1069
1070
    @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
1071
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1072
1073
1074
1075
1076
1077
1078
1079
    ) -> 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":
1080
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1081
1082
            # if last message is user, append to it to avoid two user messages in a row
            else:
1083
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1084
1085
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1086
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1087
1088
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1089

lintangsutawika's avatar
lintangsutawika committed
1090
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1091
1092
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1093
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1094
1095
1096
1097
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1098
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1099
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1100
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1101
1102
1103
1104
1105
1106
1107
        """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
1108
1109
1110
1111
1112
1113
        :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.
1114
1115
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1116
1117
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1118
1119
1120
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1121
1122
1123
1124
1125
1126
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1127
1128
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1129

KonradSzafer's avatar
KonradSzafer committed
1130
1131
1132
1133
1134
1135
1136
1137
1138
        # 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
1139
        else:
KonradSzafer's avatar
KonradSzafer committed
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
            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
1153
1154
1155
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1156
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1157
1158
1159
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1160
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1161
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1162
                )
lintangsutawika's avatar
lintangsutawika committed
1163
1164

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1165
1166
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1167
                # TODO: append prefill?
1168
1169
                if not labeled_examples:
                    return ""
1170
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1171
1172
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1173
1174
1175
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1176
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1177
1178
1179
1180
1181
1182
1183
                )
            # 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
1184
1185
1186
1187
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1188
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1189
1190
1191
1192
1193
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1194
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1195
1196
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1197
1198
1199
1200
1201
1202
                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
1203
1204
1205
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1206
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1207
1208
1209
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1210
1211
1212
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1213
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1214
1215
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1216
1217
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1218
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1219
            )
1220
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1221
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1222
1223
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1224
1225
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1226
1227
            if self.multiple_input:
                return labeled_examples
1228
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1229
                return labeled_examples + example + prefix
1230
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1231
                return [labeled_examples + ex + prefix for ex in example]
1232
1233
1234
            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
1235
                    return labeled_examples + choices[example] + prefix
1236
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1237
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1238

Baber Abbasi's avatar
Baber Abbasi committed
1239
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1240
        """Iterates over FilterEnsembles and applies them to instances"""
1241
1242
        if hasattr(self, "_filters"):
            for f in self._filters:
1243
                f.apply(self._instances)
1244
1245
1246
1247
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1248
    def should_decontaminate(self):
1249
        return self.config.should_decontaminate
1250

Baber Abbasi's avatar
Baber Abbasi committed
1251
    def doc_to_decontamination_query(self, doc: dict):
1252
        if self.config.should_decontaminate:
1253
1254
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1255
            else:
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
                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
                        )
                    )
1267

1268
    def _process_doc(self, doc: dict) -> dict:
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
        """
        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
1279
    def doc_to_text(self, doc, doc_to_text=None):
1280
1281
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1282
1283
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1284
        else:
1285
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1286

1287
        if isinstance(doc_to_text, int):
1288
            return doc_to_text
1289
        elif isinstance(doc_to_text, str):
1290
            if doc_to_text in self.features:
1291
                # if self.config.doc_to_choice is not None:
1292
1293
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1294
1295
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1296
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1297
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1298
1299
1300
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1301
        elif callable(doc_to_text):
1302
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1303
        # Used when applying a Promptsource template
1304
        elif hasattr(doc_to_text, "apply"):
1305
1306
1307
1308
1309
            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")
1310
                return self.config.fewshot_delimiter
1311
        else:
1312
            print(type(doc_to_text))
1313
            raise TypeError
1314

Yu Shi Jie's avatar
Yu Shi Jie committed
1315
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1316
1317
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1318
1319
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1320
        else:
1321
            doc_to_target = self.config.doc_to_target
1322

1323
        if isinstance(doc_to_target, int):
1324
            return doc_to_target
1325
        elif isinstance(doc_to_target, str):
1326
            if doc_to_target in self.features:
1327
                # if self.config.doc_to_choice is not None:
1328
1329
1330
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1331
            else:
lintangsutawika's avatar
lintangsutawika committed
1332
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1333
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1334
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1335
1336
1337
1338
1339
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1340
1341
1342
1343
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1344
1345
                else:
                    return target_string
1346
        elif isinstance(doc_to_target, list):
1347
            return doc_to_target
1348
        elif callable(doc_to_target):
1349
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1350
        # Used when applying a Promptsource template
1351
        elif hasattr(doc_to_target, "apply"):
1352
            applied_prompt = doc_to_target.apply(doc)
1353
1354
1355
1356
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1357
                return self.config.fewshot_delimiter
1358
1359
        else:
            raise TypeError
1360

Yu Shi Jie's avatar
Yu Shi Jie committed
1361
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1362
1363
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1364
1365
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1366
        elif self.config.doc_to_choice is None:
1367
1368
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1369
            doc_to_choice = self.config.doc_to_choice
1370

1371
        if isinstance(doc_to_choice, str):
1372
1373
1374
1375
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1376
        elif isinstance(doc_to_choice, list):
1377
            return doc_to_choice
1378
        elif isinstance(doc_to_choice, dict):
1379
1380
1381
1382
1383
1384
1385
            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
1386

1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list]:
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list]:
        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 Abbasi's avatar
Baber Abbasi committed
1433
1434
1435
1436
1437
1438
1439
1440
    def doc_to_prefix(self, doc):
        if (gen_prefix := self.config.gen_prefix) is not None:
            if gen_prefix in self.features:
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1441
1442
1443
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1444
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1445
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1446

1447
1448
        aux_arguments = None

1449
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1450
            arguments = (ctx, self.doc_to_target(doc))
1451
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1452
            arguments = (self.doc_to_target(doc),)
1453
        elif self.OUTPUT_TYPE == "multiple_choice":
1454
            choices = self.doc_to_choice(doc)
1455
            target_delimiter = self.config.target_delimiter
1456
1457
            if apply_chat_template:
                target_delimiter = ""
1458
1459
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1460
                # apply chat_template to choices if apply_chat_template
1461
                cont = self.doc_to_target(doc)
1462

1463
                arguments = [
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1474
                ]
1475
            else:
1476
                # Otherwise they are placed in the continuation
1477
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1478

1479
1480
1481
1482
1483
1484
1485
1486
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # log(P(choice|ctx) / P(choice)) = log(P(choice|ctx)) - log(P(choice))
                # in other words normalizing by subtracting the unconditional logprob of each choice.
1487
1488
1489
1490
                # 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
                ]
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505

                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)},
            }

1506
1507
1508
1509
1510
1511
1512
1513
        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)},
            }

1514
1515
1516
1517
1518
1519
1520
        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":
1521
            request_list = [
1522
1523
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1524
                    doc=doc,
1525
                    arguments=arg,
1526
                    idx=i,
1527
1528
                    **kwargs,
                )
1529
                for i, arg in enumerate(arguments)
1530
            ]
1531
1532

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1533

lintangsutawika's avatar
lintangsutawika committed
1534
        return Instance(
1535
1536
1537
1538
1539
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1540
        )
1541
1542

    def process_results(self, doc, results):
1543
1544
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1545

1546
        result_dict = {}
1547
        use_metric = list(self._metric_fn_list.keys())
1548
1549
1550
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1551
1552
1553
1554
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1555
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1556
            (loglikelihood,) = results
1557
1558
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1559
            return {
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
                **(
                    {"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
1575
            }
1576
        elif self.OUTPUT_TYPE == "multiple_choice":
1577
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1578

1579
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1580
            choices = self.doc_to_choice(doc)
1581
1582
            completion_len = np.array([float(len(i)) for i in choices])

1583
1584
            if (
                2 * len(choices) == len(lls)
1585
                and "acc_mutual_info" in self._metric_fn_list.keys()
1586
1587
1588
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1589
1590
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1591
1592
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1593
                # and this stores our "regular" conditional loglikelihoods
1594
                lls = lls[: len(choices)]
1595

1596
1597
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1598

1599
1600
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1601
            else:
1602
                gold = self.doc_to_target(doc)
1603
1604

            gold_index_error = False
1605
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1606
1607
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1608
1609
                    gold_index_error = True
            else:
1610
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1611
                    gold = gold if gold < len(choices) else -100
1612
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1613
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1614

Lintang Sutawika's avatar
Lintang Sutawika committed
1615
                if gold == -100:
1616
1617
1618
1619
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1620
                    f"Label index was not in within range of available choices,"
1621
1622
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1623

1624
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1625
1626
                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
1627
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1628
1629
1630
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1631
                # 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
1632
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1633

Lintang Sutawika's avatar
Lintang Sutawika committed
1634
1635
1636
1637
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1638
            result_dict = {
1639
                **({"acc": acc} if "acc" in use_metric else {}),
1640
1641
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1642
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1643
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1644
1645
1646
1647
1648
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1649
1650
            }

1651
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1652
1653
1654
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1655
1656
1657
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1658
        elif self.OUTPUT_TYPE == "generate_until":
1659
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1660
            result = results[0]
1661
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1662
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1663
                # it assumes that doc_to_target returns a number.
1664
1665
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1666
1667
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1668
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1669
1670
1671
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
1672
            ):
Chris's avatar
Chris committed
1673
1674
                # cast gold to the same type as result
                gold = type(result)(gold)
1675

lintangsutawika's avatar
lintangsutawika committed
1676
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1677
1678
1679
1680
1681
                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
1682
1683
1684
1685
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1686
1687
1688
1689
1690
1691
1692
1693
                    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
1694
                    else:
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
                        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
1716
                else:
1717
                    try:
1718
                        result_score = self._metric_fn_list[metric](
1719
1720
                            references=[gold],
                            predictions=[result],
1721
                            **self._metric_fn_kwargs[metric],
1722
                        )
1723
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1724
                        result_score = self._metric_fn_list[metric]([gold, result])
1725
1726
1727
1728
1729
1730
1731
                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
1732
        else:
lintangsutawika's avatar
lintangsutawika committed
1733
1734
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1735
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1736
            )
1737
1738
1739

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1740
    def aggregation(self) -> dict:
1741
1742
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1743
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1744
        return self._higher_is_better
1745

Baber Abbasi's avatar
Baber Abbasi committed
1746
1747
1748
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1749
1750
1751
1752
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1753
1754
1755
1756
1757
    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
1758
            f"num_samples={len(self.eval_docs)})"
1759
1760
        )

1761
1762

class MultipleChoiceTask(Task):
1763
    OUTPUT_TYPE = "loglikelihood"
1764

baberabb's avatar
baberabb committed
1765
    def doc_to_target(self, doc: dict) -> str:
1766
1767
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1768
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1769
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1770
1771
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1772
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1773
                doc=doc,
Baber's avatar
Baber committed
1774
                arguments=(ctx, f" {choice}"),
1775
                idx=i,
1776
1777
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1778
1779
            for i, choice in enumerate(doc["choices"])
        ]
1780

1781
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1782
1783
1784
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
        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
1796
    def higher_is_better(self) -> dict:
1797
1798
1799
1800
1801
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1802
    def aggregation(self) -> dict:
1803
1804
1805
1806
1807
1808
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1809
class PerplexityTask(Task):
1810
1811
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1812
    def has_training_docs(self) -> bool:
1813
1814
        return False

baberabb's avatar
baberabb committed
1815
    def fewshot_examples(self, k: int, rnd) -> List:
1816
1817
1818
1819
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1820
1821
        return []

baberabb's avatar
baberabb committed
1822
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1823
1824
1825
1826
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1827
1828
1829

        return ""

baberabb's avatar
baberabb committed
1830
    def higher_is_better(self) -> dict:
1831
1832
1833
1834
1835
1836
1837
1838
1839
        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
1840
    def doc_to_text(self, doc) -> str:
1841
1842
1843
1844
1845
        return ""

    def doc_to_target(self, doc):
        return doc

1846
1847
1848
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1849

lintangsutawika's avatar
lintangsutawika committed
1850
1851
1852
1853
1854
1855
1856
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1857

1858
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1859
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1860
1861
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1862
1863
1864
1865
1866
1867
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1868
    def aggregation(self) -> dict:
1869
1870
1871
1872
1873
1874
1875
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1876
    def count_bytes(cls, doc) -> int:
1877
1878
1879
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1880
    def count_words(cls, doc) -> int:
1881
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1882
        return len(re.split(r"\s+", doc))