task.py 73 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
8
from dataclasses import asdict, dataclass
9
from inspect import getsource
10
11
12
13
14
15
16
17
18
19
20
21
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
22
23
24

import datasets
import numpy as np
25
from tqdm import tqdm
26
27

from lm_eval import utils
28
from lm_eval.api import samplers
29
30
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
lintangsutawika's avatar
lintangsutawika committed
31
from lm_eval.api.registry import (
32
33
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
34
    get_aggregation,
35
    get_metric,
36
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
38
)
39
from lm_eval.caching.cache import load_from_cache, save_to_cache
40
41
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt
42
from lm_eval.utils import validate_index
43

44

45
46
47
48
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
49
    "generate_until",
50
51
]

Lintang Sutawika's avatar
Lintang Sutawika committed
52
eval_logger = logging.getLogger(__name__)
53

lintangsutawika's avatar
lintangsutawika committed
54

55
56
@dataclass
class TaskConfig(dict):
57
    # task naming/registry
58
59
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
60
    tag: Optional[Union[str, list]] = None
61
62
63
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
64
    custom_dataset: Optional[Callable] = None
65
66
67
68
69
70
    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
71
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
72
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
73
    )
74
75
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
76
77
78
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
79
    doc_to_image: Union[Callable, str] = None
80
    doc_to_audio: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
81
    unsafe_code: bool = False
82
83
84
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
85
    description: str = ""
86
87
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
88
    fewshot_config: Optional[dict] = None
89
    # runtime configuration options
90
    num_fewshot: Optional[int] = None
91
    # scoring options
92
93
94
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
95
    repeats: int = 1
96
    filter_list: Optional[Union[str, list]] = None
97
    should_decontaminate: bool = False
98
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
99
    gen_prefix: Optional[str] = None
100
101
    multiple_inputs: bool = False
    multiple_targets: bool = False
102
103
104
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
105

Ethan Smith's avatar
Ethan Smith committed
106
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
107
        if self.generation_kwargs is not None:
108
            if self.output_type != "generate_until":
109
                eval_logger.warning(
110
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
111
112
113
114
115
116
117
118
                )

            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
119
120
121
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
122
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
123
        else:
124
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
125
126
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
127
128
129
130
131
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
132
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
133
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
134
                }
Baber Abbasi's avatar
Baber Abbasi committed
135
136
137
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
138

139
140
141
    def __getitem__(self, item):
        return getattr(self, item)

142
143
144
    def __setitem__(self, item, value):
        return setattr(self, item, value)

145
    def to_dict(self, keep_callable: bool = False) -> dict:
146
147
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
148
        Used for dumping results alongside full task configuration
149

haileyschoelkopf's avatar
haileyschoelkopf committed
150
151
152
153
154
155
156
157
158
159
        :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)
160
161
162
163
164
165
166
167
168
169
            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
170
        return cfg_dict
171

172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
    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)

188
189
190
191
192
193
194
195
196
197
198

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)
    """

199
    VERSION: Optional[Union[int, str]] = None
200

201
202
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
203
    DATASET_PATH: Optional[str] = None
204
205

    # The name of a subset within `DATASET_PATH`.
206
    DATASET_NAME: Optional[str] = None
207

208
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
209

210
211
    def __init__(
        self,
212
213
214
215
        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
216
    ) -> None:
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
        """
        :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)
239
240
241
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
242

243
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
244

lintangsutawika's avatar
lintangsutawika committed
245
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
246
247
248
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
249

250
251
252
253
254
255
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
        """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.
        """
280
281
282
283
284
285
286
        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,
        )
287

288
    @property
289
    def config(self) -> TaskConfig:
290
291
292
        """Returns the TaskConfig associated with this class."""
        return self._config

293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
    @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

308
    def training_docs(self) -> Iterable:
309
310
311
312
313
314
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

315
    def validation_docs(self) -> Iterable:
316
317
318
319
320
321
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

322
    def test_docs(self) -> Iterable:
323
324
325
326
327
328
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

329
    def fewshot_docs(self) -> Iterable:
330
331
332
333
334
335
336
337
338
        """
        :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
339
340
341
342
343
            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."
                )
344
345
            return self.test_docs()

346
    def _process_doc(self, doc: dict) -> dict:
347
348
349
350
351
352
353
354
355
        """
        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
356

357
    @property
358
    def instances(self) -> List[Instance]:
359
360
361
362
363
364
365
366
367
368
369
        """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)

370
371
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
372
373
374
375
376
377
378
379
380
381
382
            "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

383
384
385
386
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

387
388
389
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
390
391
392
    def doc_to_prefix(self, doc):
        return ""

393
394
    def build_all_requests(
        self,
395
        *,
396
        limit: Union[int, None] = None,
397
        samples: Optional[List[int]] = None,
398
399
400
401
402
403
404
405
406
        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 = "",
407
    ) -> None:
408
        """Build a set of Instances for a task, and store them in task.instances"""
409
410
411
412

        # used with caching
        og_limit = limit

413
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
414
415
416
417
418
419
420
        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 ""
        )
421
        cache_key += f"-tokenizer{tokenizer_name}"
422

Baber Abbasi's avatar
Baber Abbasi committed
423
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
424
425
426
427
428
429
430
431
432
433
434
435
436

        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
437
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
438

439
        instances = []
440
441
442
443
444
445
446
447
448
449

        # 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(
450
451
452
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
453
454
455
456
457
458
459
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
460
        ):
461
            # sample fewshot context #TODO: need to offset doc_id by rank now!
462
            fewshot_ctx = self.fewshot_context(
463
                doc,
464
465
466
467
468
469
470
                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
471
                gen_prefix=self.doc_to_prefix(doc),
472
            )
473

474
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
475
476
477
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
478
                metadata=(self.config["task"], doc_id, self.config.repeats),
479
                apply_chat_template=apply_chat_template,
480
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
481
            )
482
483
484
485

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

486
487
488
489
490
491
492
493
494
495
496
497
498
            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
499

500
501
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
502

503
504
505
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
    @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
522
            The number of times each instance in a dataset is inferred on. Defaults to 1,
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
552
553
554
555
556
557
            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

558
559
560
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
561
562
563
564
565
566
567
568
569
570
    @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))

571
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
572
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
573
574
575
576
577
578
579
        """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
580
581
582
583
584
        :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.
585
586
587
        :returns: str
            The fewshot context.
        """
588
        if rnd is None:
589
590
591
592
593
594
            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
595

596
        description = description if description else ""
597
598

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
599
            labeled_examples = ""
600
        else:
lintangsutawika's avatar
lintangsutawika committed
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
            # 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
625
            )
626
627

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
628
        return description + labeled_examples + example
629

630
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
631
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
632
633
        if hasattr(self, "_filters"):
            for f in self._filters:
634
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
635
636
637
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
638

baberabb's avatar
baberabb committed
639
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
640
        """Returns the config as a dictionary."""
641
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
642
        # (num_fewshot)
643
        return self.config.to_dict()
644

Baber Abbasi's avatar
Baber Abbasi committed
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
679
680
681
682
683
684
    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)

685
686
687
688
689
    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

690
691
692
693
694
695
696
    @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:
697
698
699
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
700
701

    def doc_iterator(
702
703
704
705
706
707
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
708
    ) -> Iterator[Tuple[int, Any]]:
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
        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),
            )
731
732
        return doc_iterator

733
734

class ConfigurableTask(Task):
735
    VERSION = "Yaml"
736
    OUTPUT_TYPE = None
737
    CONFIG = None
738
739

    def __init__(
740
741
742
743
744
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
745
    ) -> None:  # TODO no super() call here
746
        # Get pre-configured attributes
747
        self._config = self.CONFIG
748

749
        # Use new configurations if there was no preconfiguration
750
        if self.config is None:
751
            self._config = TaskConfig(**config)
752
753
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
754
            if config is not None:
755
                self._config.__dict__.update(config)
756

757
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
758
759
760
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
761

762
763
764
765
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

766
        if self.config.output_type is not None:
767
768
769
770
            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)}'"
                )
771
            self.OUTPUT_TYPE = self.config.output_type
772

773
774
775
776
777
778
        self.multiple_targets = self.config.multiple_targets
        self.multiple_inputs = self.config.multiple_inputs
        assert not (self.multiple_targets and self.multiple_inputs), (
            "Cannot have both multiple_targets and multiple_inputs"
        )

779
780
781
782
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

783
784
785
786
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
787
788
789
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

790
791
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
792

793
794
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
795

796
797
798
799
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
800

801
        if self.config.metric_list is None:
802
            # TODO: handle this in TaskConfig.__post_init__ ?
803
804
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

805
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
806
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
807
                self._metric_fn_kwargs[metric_name] = {}
808
809
810
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
811
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
812
        else:
813
            for metric_config in self.config.metric_list:
814
815
816
817
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
818
819
820
821
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
822
823
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
824
                }
Chris's avatar
Chris committed
825
826
827
828
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
829

830
                if self.config.process_results is not None:
831
832
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
833
834
835
836
837
838
                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
839
840
841
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
842
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
843

844
                if "aggregation" in metric_config:
845
                    agg_name = metric_config["aggregation"]
846
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
847
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
848
                    elif callable(agg_name):  # noqa: E721
849
850
851
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
852
                else:
853
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
854
                    metric_agg = get_metric_aggregation(metric_name)
855
                    eval_logger.warning(
856
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
857
858
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
859
                    )
860
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
861

862
863
864
865
866
867
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
868
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
869
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
870
                        f"higher_is_better={is_higher_better(metric_name)}"
871
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
872
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
873

874
        self.download(self.config.dataset_kwargs)
875
876
877
        self._training_docs = None
        self._fewshot_docs = None

878
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
879
            self._filters = []
880
            for filter_config in self.config.filter_list:
881
882
883
884
885
886
887
888
889
                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
890
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
891
        else:
Baber Abbasi's avatar
Baber Abbasi committed
892
893
894
895
            # 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."
            )
896
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
897

898
899
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
900
            self.prompt = get_prompt(
901
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
902
            )
903
904
905
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
906
        if self.fewshot_docs() is not None:
907
908
909
910
            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
911
912
913
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
            )
            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)}"
                )
930

931
        self.task_docs = self.eval_docs
932

933
        # Test One Doc
934
935
        self.features = list(self.task_docs.features.keys())
        test_doc = self.task_docs[0]
936
        test_text = self.doc_to_text(test_doc)
937
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
938

939
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
940
            test_choice = self.doc_to_choice(test_doc)
941
942
943
944
945
946
947
948
            if self.multiple_inputs:
                # we require:
                # doc_to_text: int
                # doc_to_choice: list
                # doc_to_target: str
                # e.g. text: 1, choice: [Maria was better than Sarah, Sarah was better than Sarah]
                # target: so she was envious
                assert isinstance(test_text, int), (
Baber's avatar
Baber committed
949
                    f"[{self.config.task}] doc_to_text must return int for multiple inputs"
950
951
                )
                assert isinstance(test_target, str), (
Baber's avatar
Baber committed
952
                    f"[{self.config.task}] doc_to_target must return str for multiple inputs"
953
954
                )
                assert self.config.output_type != "generate_until", (
Baber's avatar
Baber committed
955
                    f"[{self.config.task}] Only multiple-choice tasks can be used with multiple inputs"
956
957
958
959
960
961
962
963
964
                )
                test_text = test_choice[0]

            elif self.multiple_targets:
                # we require:
                # doc_to_text: str
                # doc_to_choice: list
                # doc_to_target: list
                assert isinstance(test_target, (list, tuple)), (
Baber's avatar
Baber committed
965
                    f"[{self.config.task}] doc_to_target must be an iterable for multiple targets"
966
967
                )
                test_target = test_target[0]
968
            else:
969
                assert isinstance(test_target, int), (
Baber's avatar
Baber committed
970
                    f"[{self.config.task}] doc_to_target must return int for multiple-choice tasks"
Baber Abbasi's avatar
Baber Abbasi committed
971
                )
lintangsutawika's avatar
lintangsutawika committed
972
                test_target = test_choice[test_target]
973

974
975
            assert hasattr(test_choice, "__iter__") and not isinstance(
                test_choice, (str, bytes)
Baber's avatar
Baber committed
976
            ), f"[{self.config.task}] doc_to_choice must be an iterable!"
977
978
979

            for choice in test_choice:
                choice_has_whitespace = choice[0].isspace()
980
                delimiter_has_whitespace = (
981
                    self.config.target_delimiter.rstrip()
982
                    != self.config.target_delimiter
983
                )
984

985
                if delimiter_has_whitespace and choice_has_whitespace:
986
                    eval_logger.debug(
Baber's avatar
Baber committed
987
988
                        f'[{self.config.task}] Both target_delimiter "{self.config.target_delimiter}" and target '
                        f'choice: "{choice}" have whitespace'
989
990
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
991
                    eval_logger.debug(
Baber's avatar
Baber committed
992
993
                        f'[{self.config.task}] Both target_delimiter "{self.config.target_delimiter}" and target '
                        f'choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
994
995
                    )

Baber Abbasi's avatar
Baber Abbasi committed
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
        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 {},
            )
1013

baberabb's avatar
baberabb committed
1014
    def has_training_docs(self) -> bool:
1015
        if self.config.training_split is not None:
1016
1017
1018
1019
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1020
    def has_validation_docs(self) -> bool:
1021
        if self.config.validation_split is not None:
1022
1023
1024
1025
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1026
    def has_test_docs(self) -> bool:
1027
        if self.config.test_split is not None:
1028
1029
1030
1031
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1032
    def training_docs(self) -> datasets.Dataset:
1033
        if self.has_training_docs():
1034
1035
1036
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1037
                )
1038
            return self.dataset[self.config.training_split]
1039

baberabb's avatar
baberabb committed
1040
    def validation_docs(self) -> datasets.Dataset:
1041
        if self.has_validation_docs():
1042
1043
1044
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1045
                )
1046
            return self.dataset[self.config.validation_split]
1047

baberabb's avatar
baberabb committed
1048
    def test_docs(self) -> datasets.Dataset:
1049
        if self.has_test_docs():
1050
1051
1052
            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]
1053

1054
    def fewshot_docs(self):
1055
        if self.config.fewshot_split is not None:
1056
1057
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1058
            return self.dataset[self.config.fewshot_split]
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
        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."
                )
1071
        else:
1072
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1073
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1074
                    f"[Task: {self.config.task}] "
1075
1076
1077
1078
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1079

KonradSzafer's avatar
KonradSzafer committed
1080
1081
1082
1083
1084
    @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
1085
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1086
1087
1088
1089
1090
1091
1092
1093
    ) -> 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":
1094
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1095
1096
            # if last message is user, append to it to avoid two user messages in a row
            else:
1097
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1098
1099
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1100
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1101
1102
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1103

lintangsutawika's avatar
lintangsutawika committed
1104
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1105
1106
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1107
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1108
1109
1110
1111
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1112
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1113
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1114
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1115
1116
1117
1118
1119
1120
1121
        """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
1122
1123
1124
1125
1126
1127
        :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.
1128
1129
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1130
1131
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1132
1133
1134
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1135
1136
1137
1138
1139
1140
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1141
1142
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1143

KonradSzafer's avatar
KonradSzafer committed
1144
1145
1146
1147
1148
1149
1150
1151
1152
        # 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
1153
        else:
KonradSzafer's avatar
KonradSzafer committed
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
            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
1167
1168
1169
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1170
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1171
1172
1173
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1174
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1175
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1176
                )
lintangsutawika's avatar
lintangsutawika committed
1177
1178

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1179
        if apply_chat_template:
1180
            if self.multiple_inputs:
Baber Abbasi's avatar
Baber Abbasi committed
1181
                # TODO: append prefill?
1182
1183
                if not labeled_examples:
                    return ""
1184
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1185
1186
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1187
1188
1189
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1190
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1191
1192
1193
1194
1195
1196
1197
                )
            # 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
1198
1199
1200
1201
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1202
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1203
1204
1205
1206
1207
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1208
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1209
1210
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1211
1212
1213
1214
1215
1216
                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
1217
1218
1219
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1220
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1221
1222
1223
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1224
1225
1226
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1227
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1228
1229
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1230
1231
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1232
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1233
            )
1234
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1235
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1236
1237
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1238
1239
                else ""
            )
1240
            if self.multiple_inputs:
KonradSzafer's avatar
KonradSzafer committed
1241
                return labeled_examples
1242
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1243
                return labeled_examples + example + prefix
1244
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1245
                return [labeled_examples + ex + prefix for ex in example]
1246
1247
1248
            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
1249
                    return labeled_examples + choices[example] + prefix
1250
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1251
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1252

Baber Abbasi's avatar
Baber Abbasi committed
1253
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1254
        """Iterates over FilterEnsembles and applies them to instances"""
1255
1256
        if hasattr(self, "_filters"):
            for f in self._filters:
1257
                f.apply(self._instances)
1258
1259
1260
1261
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1262
    def should_decontaminate(self):
1263
        return self.config.should_decontaminate
1264

Baber Abbasi's avatar
Baber Abbasi committed
1265
    def doc_to_decontamination_query(self, doc: dict):
1266
        if self.config.should_decontaminate:
1267
1268
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1269
            else:
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
                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
                        )
                    )
1281

1282
    def _process_doc(self, doc: dict) -> dict:
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
        """
        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
1293
    def doc_to_text(self, doc, doc_to_text=None):
1294
1295
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1296
1297
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1298
        else:
1299
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1300

1301
        if isinstance(doc_to_text, int):
1302
            return doc_to_text
1303
        elif isinstance(doc_to_text, str):
1304
            if doc_to_text in self.features:
1305
                # if self.config.doc_to_choice is not None:
1306
1307
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1308
1309
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1310
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1311
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1312
1313
1314
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1315
        elif callable(doc_to_text):
1316
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1317
        # Used when applying a Promptsource template
1318
        elif hasattr(doc_to_text, "apply"):
1319
1320
1321
1322
1323
            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")
1324
                return self.config.fewshot_delimiter
1325
        else:
1326
            print(type(doc_to_text))
1327
            raise TypeError
1328

Yu Shi Jie's avatar
Yu Shi Jie committed
1329
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1330
1331
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1332
1333
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1334
        else:
1335
            doc_to_target = self.config.doc_to_target
1336

1337
        if isinstance(doc_to_target, int):
1338
            return doc_to_target
1339
        elif isinstance(doc_to_target, str):
1340
            if doc_to_target in self.features:
1341
                # if self.config.doc_to_choice is not None:
1342
1343
1344
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1345
            else:
lintangsutawika's avatar
lintangsutawika committed
1346
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1347
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1348
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1349
1350
1351
1352
1353
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1354
1355
1356
1357
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1358
1359
                else:
                    return target_string
1360
        elif isinstance(doc_to_target, list):
1361
            return doc_to_target
1362
        elif callable(doc_to_target):
1363
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1364
        # Used when applying a Promptsource template
1365
        elif hasattr(doc_to_target, "apply"):
1366
            applied_prompt = doc_to_target.apply(doc)
1367
1368
1369
1370
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1371
                return self.config.fewshot_delimiter
1372
1373
        else:
            raise TypeError
1374

Yu Shi Jie's avatar
Yu Shi Jie committed
1375
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1376
1377
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1378
1379
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1380
        elif self.config.doc_to_choice is None:
1381
1382
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1383
            doc_to_choice = self.config.doc_to_choice
1384

1385
        if isinstance(doc_to_choice, str):
1386
1387
1388
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
1389
                return utils.apply_template(doc_to_choice, doc)
1390
        elif isinstance(doc_to_choice, list):
1391
            return doc_to_choice
1392
        elif isinstance(doc_to_choice, dict):
1393
1394
1395
1396
1397
1398
1399
            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
1400

1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
    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

1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
    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
1447
1448
1449
1450
1451
1452
1453
1454
    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
1455
1456
1457
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1458
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1459
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1460

1461
1462
        aux_arguments = None

1463
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1464
            arguments = (ctx, self.doc_to_target(doc))
1465
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1466
            arguments = (self.doc_to_target(doc),)
1467
        elif self.OUTPUT_TYPE == "multiple_choice":
1468
            choices = self.doc_to_choice(doc)
1469
            target_delimiter = self.config.target_delimiter
1470
1471
            if apply_chat_template:
                target_delimiter = ""
1472
            if self.multiple_inputs:
1473
                # If there are multiple inputs, choices are placed in the ctx
1474
                # apply chat_template to choices if apply_chat_template
1475
                cont = self.doc_to_target(doc)
1476

1477
                arguments = [
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1488
                ]
1489
            else:
1490
                # Otherwise they are placed in the continuation
1491
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1492

1493
1494
1495
1496
1497
1498
1499
1500
            # 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.
1501
1502
1503
1504
                # 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
                ]
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519

                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)},
            }

1520
1521
1522
1523
1524
1525
1526
1527
        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)},
            }

1528
1529
1530
1531
1532
1533
1534
        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":
1535
            request_list = [
1536
1537
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1538
                    doc=doc,
1539
                    arguments=arg,
1540
                    idx=i,
1541
1542
                    **kwargs,
                )
1543
                for i, arg in enumerate(arguments)
1544
            ]
1545
1546

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1547

lintangsutawika's avatar
lintangsutawika committed
1548
        return Instance(
1549
1550
1551
1552
1553
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1554
        )
1555
1556

    def process_results(self, doc, results):
1557
1558
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1559

1560
        result_dict = {}
1561
        use_metric = list(self._metric_fn_list.keys())
1562
1563
1564
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1565
1566
1567
1568
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1569
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1570
            (loglikelihood,) = results
1571
1572
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1573
            return {
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
                **(
                    {"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
1589
            }
1590
        elif self.OUTPUT_TYPE == "multiple_choice":
1591
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1592

1593
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1594
            choices = self.doc_to_choice(doc)
1595
1596
            completion_len = np.array([float(len(i)) for i in choices])

1597
1598
            if (
                2 * len(choices) == len(lls)
1599
                and "acc_mutual_info" in self._metric_fn_list.keys()
1600
1601
1602
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1603
1604
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1605
1606
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1607
                # and this stores our "regular" conditional loglikelihoods
1608
                lls = lls[: len(choices)]
1609

1610
1611
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1612

1613
1614
1615
1616
1617
            gold = (
                self.doc_to_text(doc)
                if not self.multiple_inputs
                else self.doc_to_text(doc)
            )
1618

1619
            if isinstance(gold, list):
1620
1621
                gold = [validate_index(g, len(choices)) for g in gold]
                gold_index_error = -100 in gold
1622
            else:
1623
                if isinstance(gold, int):
1624
                    gold = validate_index(gold, len(choices))
1625
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1626
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1627

1628
                gold_index_error = gold == -100
1629
1630
1631

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1632
                    f"Label index was not in within range of available choices,"
1633
1634
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1635

1636
            if self.multiple_targets:
lintangsutawika's avatar
lintangsutawika committed
1637
1638
                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
1639
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1640
1641
1642
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1643
                # 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
1644
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1645

Lintang Sutawika's avatar
Lintang Sutawika committed
1646
1647
1648
1649
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1650
            result_dict = {
1651
                **({"acc": acc} if "acc" in use_metric else {}),
1652
1653
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1654
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1655
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1656
1657
1658
1659
1660
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1661
1662
            }

1663
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1664
1665
1666
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1667
1668
1669
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1670
        elif self.OUTPUT_TYPE == "generate_until":
1671
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1672
            result = results[0]
1673
1674
1675
1676
1677
1678
1679
1680
1681
1682
            # we expect multiple_targets to be a list.
            if self.multiple_target:
                gold = list(gold)
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
            ):
                # cast gold to the same type as result
                gold = type(result)(gold)

lintangsutawika's avatar
lintangsutawika committed
1683
            for metric in self._metric_fn_list.keys():
1684
1685
1686
1687
1688
1689
1690
1691
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
                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 = []
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
                    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
                    else:
                        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
                else:
                    try:
                        result_score = self._metric_fn_list[metric](
                            references=[gold],
                            predictions=[result],
                            **self._metric_fn_kwargs[metric],
                        )
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
                        result_score = self._metric_fn_list[metric]([gold, result])
1732
1733
1734
1735
1736
1737
1738
                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
1739
        else:
lintangsutawika's avatar
lintangsutawika committed
1740
1741
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1742
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1743
            )
1744
1745
1746

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1747
    def aggregation(self) -> dict:
1748
1749
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1750
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1751
        return self._higher_is_better
1752

Baber Abbasi's avatar
Baber Abbasi committed
1753
1754
1755
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1756
1757
1758
1759
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1760
1761
1762
1763
1764
    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
1765
            f"num_samples={len(self.eval_docs)})"
1766
1767
        )

1768
1769

class MultipleChoiceTask(Task):
1770
    OUTPUT_TYPE = "loglikelihood"
1771

baberabb's avatar
baberabb committed
1772
    def doc_to_target(self, doc: dict) -> str:
1773
1774
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1775
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1776
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1777
1778
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1779
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1780
                doc=doc,
1781
                arguments=(ctx, " {}".format(choice)),
1782
                idx=i,
1783
1784
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1785
1786
            for i, choice in enumerate(doc["choices"])
        ]
1787

1788
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1789
1790
1791
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1792
1793
1794
1795
1796
1797
1798
1799
1800
1801
1802
        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
1803
    def higher_is_better(self) -> dict:
1804
1805
1806
1807
1808
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1809
    def aggregation(self) -> dict:
1810
1811
1812
1813
1814
1815
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1816
class PerplexityTask(Task):
1817
1818
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1819
    def has_training_docs(self) -> bool:
1820
1821
        return False

baberabb's avatar
baberabb committed
1822
    def fewshot_examples(self, k: int, rnd) -> List:
1823
1824
1825
1826
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1827
1828
        return []

baberabb's avatar
baberabb committed
1829
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1830
1831
1832
1833
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1834
1835
1836

        return ""

baberabb's avatar
baberabb committed
1837
    def higher_is_better(self) -> dict:
1838
1839
1840
1841
1842
1843
1844
1845
1846
        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
1847
    def doc_to_text(self, doc) -> str:
1848
1849
1850
1851
1852
        return ""

    def doc_to_target(self, doc):
        return doc

1853
1854
1855
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1856

lintangsutawika's avatar
lintangsutawika committed
1857
1858
1859
1860
1861
1862
1863
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1864

1865
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1866
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1867
1868
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1869
1870
1871
1872
1873
1874
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1875
    def aggregation(self) -> dict:
1876
1877
1878
1879
1880
1881
1882
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1883
    def count_bytes(cls, doc) -> int:
1884
1885
1886
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1887
    def count_words(cls, doc) -> int:
1888
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1889
        return len(re.split(r"\s+", doc))