task.py 72.6 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
8
from dataclasses import asdict, dataclass
9
from inspect import getsource
10
11
12
13
14
15
16
17
18
19
20
21
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
22
23
24

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

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

43

44
45
46
47
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
48
    "generate_until",
49
50
]

Lintang Sutawika's avatar
Lintang Sutawika committed
51
eval_logger = logging.getLogger(__name__)
52

lintangsutawika's avatar
lintangsutawika committed
53

54
55
@dataclass
class TaskConfig(dict):
56
    # task naming/registry
57
58
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
59
    tag: Optional[Union[str, list]] = None
60
61
62
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
63
    custom_dataset: Optional[Callable] = None
64
65
66
67
68
69
    dataset_path: Optional[str] = None
    dataset_name: Optional[str] = None
    dataset_kwargs: Optional[dict] = None
    training_split: Optional[str] = None
    validation_split: Optional[str] = None
    test_split: Optional[str] = None
70
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
71
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
72
    )
73
74
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
75
76
77
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
78
    doc_to_image: Union[Callable, str] = None
79
    doc_to_audio: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
80
    unsafe_code: bool = False
81
82
83
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
84
    description: str = ""
85
86
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
87
    fewshot_config: Optional[dict] = None
88
    # runtime configuration options
89
    num_fewshot: Optional[int] = None
90
    # scoring options
91
92
93
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
94
    repeats: int = 1
95
    filter_list: Optional[Union[str, list]] = None
96
    should_decontaminate: bool = False
97
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
98
    gen_prefix: Optional[str] = None
99
100
101
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
102

Ethan Smith's avatar
Ethan Smith committed
103
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
104
        if self.generation_kwargs is not None:
105
            if self.output_type != "generate_until":
106
                eval_logger.warning(
107
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
108
109
110
111
112
113
114
115
                )

            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
116
117
118
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
119
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
120
        else:
121
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
122
123
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
124
125
126
127
128
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
129
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
130
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
131
                }
Baber Abbasi's avatar
Baber Abbasi committed
132
133
134
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
135

136
137
138
    def __getitem__(self, item):
        return getattr(self, item)

139
140
141
    def __setitem__(self, item, value):
        return setattr(self, item, value)

142
    def to_dict(self, keep_callable: bool = False) -> dict:
143
144
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
145
        Used for dumping results alongside full task configuration
146

haileyschoelkopf's avatar
haileyschoelkopf committed
147
148
149
150
151
152
153
154
155
156
        :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)
157
158
159
160
161
162
163
164
165
166
            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
167
        return cfg_dict
168

169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
    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)

185
186
187
188
189
190
191
192
193
194
195

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)
    """

196
    VERSION: Optional[Union[int, str]] = None
197

198
199
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
200
    DATASET_PATH: Optional[str] = None
201
202

    # The name of a subset within `DATASET_PATH`.
203
    DATASET_NAME: Optional[str] = None
204

205
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
206

207
208
    def __init__(
        self,
209
210
211
212
        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
213
    ) -> None:
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
        """
        :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)
236
237
238
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
239

240
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
241

lintangsutawika's avatar
lintangsutawika committed
242
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
243
244
245
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
246

247
248
249
250
251
252
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
        """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.
        """
277
278
279
280
281
282
283
        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,
        )
284

285
    @property
286
    def config(self) -> TaskConfig:
287
288
289
        """Returns the TaskConfig associated with this class."""
        return self._config

290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
    @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

305
    def training_docs(self) -> Iterable:
306
307
308
309
310
311
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

312
    def validation_docs(self) -> Iterable:
313
314
315
316
317
318
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

319
    def test_docs(self) -> Iterable:
320
321
322
323
324
325
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

326
    def fewshot_docs(self) -> Iterable:
327
328
329
330
331
332
333
334
335
        """
        :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
336
337
338
339
340
            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."
                )
341
342
            return self.test_docs()

343
    def _process_doc(self, doc: dict) -> dict:
344
345
346
347
348
349
350
351
352
        """
        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
353

354
    @property
355
    def instances(self) -> List[Instance]:
356
357
358
359
360
361
362
363
364
365
366
        """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)

367
368
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
369
370
371
372
373
374
375
376
377
378
379
            "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

380
381
382
383
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

384
385
386
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
387
388
389
    def doc_to_prefix(self, doc):
        return ""

390
391
    def build_all_requests(
        self,
392
        *,
393
        limit: Union[int, None] = None,
394
        samples: Optional[List[int]] = None,
395
396
397
398
399
400
401
402
403
        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 = "",
404
        question_suffix: str = ""
405
    ) -> None:
406
        """Build a set of Instances for a task, and store them in task.instances"""
407
408
409
410

        # used with caching
        og_limit = limit

411
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
412
413
414
415
416
417
418
        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 ""
        )
419
        cache_key += f"-tokenizer{tokenizer_name}"
420

Baber Abbasi's avatar
Baber Abbasi committed
421
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
422
423
424
425
426
427
428
429
430
431
432
433
434

        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
435
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
436

437
        instances = []
438
439
440
441
442
443
444
445
446
447

        # 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(
448
449
450
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
451
452
453
454
455
456
457
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
458
        ):
459
            # sample fewshot context #TODO: need to offset doc_id by rank now!
460
            fewshot_ctx = self.fewshot_context(
461
                doc,
462
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
463
464
465
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
466
                chat_template,
Baber Abbasi's avatar
Baber Abbasi committed
467
                gen_prefix=self.doc_to_prefix(doc),
468
                question_suffix=question_suffix,
469
            )
470

471
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
472
473
474
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
475
                metadata=(self.config["task"], doc_id, self.config.repeats),
476
                apply_chat_template=apply_chat_template,
477
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
478
            )
479
480
481
482

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

483
484
485
486
487
488
489
490
491
492
493
494
495
            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
496

497
498
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
499

500
501
502
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
    @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
519
            The number of times each instance in a dataset is inferred on. Defaults to 1,
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
            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

555
556
557
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
558
559
560
561
562
563
564
565
566
567
    @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))

568
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
569
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
570
571
572
573
574
575
576
        """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
577
578
579
580
581
        :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.
582
583
584
        :returns: str
            The fewshot context.
        """
585
        if rnd is None:
586
587
588
589
590
591
            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
592

593
        description = description if description else ""
594
595

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
596
            labeled_examples = ""
597
        else:
lintangsutawika's avatar
lintangsutawika committed
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
            # 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
622
            )
623
624

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
625
        return description + labeled_examples + example
626

627
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
628
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
629
630
        if hasattr(self, "_filters"):
            for f in self._filters:
631
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
632
633
634
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
635

baberabb's avatar
baberabb committed
636
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
637
        """Returns the config as a dictionary."""
638
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
639
        # (num_fewshot)
640
        return self.config.to_dict()
641

Baber Abbasi's avatar
Baber Abbasi committed
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
    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)

682
683
684
685
686
    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

687
688
689
690
691
692
693
    @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:
694
695
696
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
697
698

    def doc_iterator(
699
700
701
702
703
704
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
705
    ) -> Iterator[Tuple[int, Any]]:
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
        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),
            )
728
729
        return doc_iterator

730
731

class ConfigurableTask(Task):
732
    VERSION = "Yaml"
733
    OUTPUT_TYPE = None
734
    CONFIG = None
735
736

    def __init__(
737
738
739
740
741
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
742
    ) -> None:  # TODO no super() call here
743
        # Get pre-configured attributes
744
        self._config = self.CONFIG
745

746
        # Use new configurations if there was no preconfiguration
747
        if self.config is None:
748
            self._config = TaskConfig(**config)
749
750
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
751
            if config is not None:
752
                self._config.__dict__.update(config)
753

754
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
755
756
757
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
758

759
760
761
762
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

763
        if self.config.output_type is not None:
764
765
766
767
            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)}'"
                )
768
            self.OUTPUT_TYPE = self.config.output_type
769

770
771
772
773
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

774
775
776
777
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
778
779
780
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

781
782
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
783

784
785
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
786

787
788
789
790
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
791

792
        if self.config.metric_list is None:
793
            # TODO: handle this in TaskConfig.__post_init__ ?
794
795
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

796
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
797
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
798
                self._metric_fn_kwargs[metric_name] = {}
799
800
801
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
802
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
803
        else:
804
            for metric_config in self.config.metric_list:
805
806
807
808
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
809
810
811
812
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
813
814
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
815
                }
Chris's avatar
Chris committed
816
817
818
819
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
820

821
                if self.config.process_results is not None:
822
823
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
824
825
826
827
828
829
                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
830
831
832
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
833
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
834

835
                if "aggregation" in metric_config:
836
                    agg_name = metric_config["aggregation"]
837
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
838
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
839
                    elif callable(agg_name):  # noqa: E721
840
841
842
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
843
                else:
844
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
845
                    metric_agg = get_metric_aggregation(metric_name)
846
                    eval_logger.warning(
847
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
848
849
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
850
                    )
851
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
852

853
854
855
856
857
858
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
859
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
860
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
861
                        f"higher_is_better={is_higher_better(metric_name)}"
862
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
863
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
864

865
        self.download(self.config.dataset_kwargs)
866
867
868
        self._training_docs = None
        self._fewshot_docs = None

869
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
870
            self._filters = []
871
            for filter_config in self.config.filter_list:
872
873
874
875
876
877
878
879
880
                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
881
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
882
        else:
Baber Abbasi's avatar
Baber Abbasi committed
883
884
885
886
            # 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."
            )
887
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
888

889
890
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
891
            self.prompt = get_prompt(
892
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
893
            )
894
895
896
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
897
        if self.fewshot_docs() is not None:
898
899
900
901
            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
902
903
904
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
            )
            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)}"
                )
921

922
        self.task_docs = self.eval_docs
923

924
        # Test One Doc
925
        self.features = list(self.task_docs.features.keys())
926
927
        self.multiple_input = 0
        self.multiple_target = 0
928
        test_doc = self.task_docs[0]
929
        test_text = self.doc_to_text(test_doc)
930
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
931

932
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
933
            test_choice = self.doc_to_choice(test_doc)
934
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
935
                eval_logger.error("doc_to_choice must return list")
936
937
            else:
                num_choice = len(test_choice)
938

939
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
940
941
942
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
943
                self.multiple_input = num_choice
944
945
        else:
            test_choice = None
946

947
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
948
949
950
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
951
            self.multiple_target = len(test_target)
952
        else:
953
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
954
                test_target = test_choice[test_target]
955
            else:
lintangsutawika's avatar
lintangsutawika committed
956
                test_target = str(test_target)
957

958
959
960
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
961
            check_choices = [test_target]
962
963
964
965
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = True if choice[0].isspace() else False
                delimiter_has_whitespace = (
966
967
                    True
                    if self.config.target_delimiter.rstrip()
968
                    != self.config.target_delimiter
969
                    else False
970
                )
971

972
                if delimiter_has_whitespace and choice_has_whitespace:
973
974
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
975
976
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
977
                    eval_logger.debug(
978
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" do not have whitespace, ignore if the language you are evaluating on does not require/use whitespace'
979
980
                    )

Baber Abbasi's avatar
Baber Abbasi committed
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
    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 {},
            )
998

baberabb's avatar
baberabb committed
999
    def has_training_docs(self) -> bool:
1000
        if self.config.training_split is not None:
1001
1002
1003
1004
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1005
    def has_validation_docs(self) -> bool:
1006
        if self.config.validation_split is not None:
1007
1008
1009
1010
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1011
    def has_test_docs(self) -> bool:
1012
        if self.config.test_split is not None:
1013
1014
1015
1016
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1017
    def training_docs(self) -> datasets.Dataset:
1018
        if self.has_training_docs():
1019
1020
1021
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1022
                )
1023
            return self.dataset[self.config.training_split]
1024

baberabb's avatar
baberabb committed
1025
    def validation_docs(self) -> datasets.Dataset:
1026
        if self.has_validation_docs():
1027
1028
1029
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1030
                )
1031
            return self.dataset[self.config.validation_split]
1032

baberabb's avatar
baberabb committed
1033
    def test_docs(self) -> datasets.Dataset:
1034
        if self.has_test_docs():
1035
1036
1037
            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]
1038

1039
    def fewshot_docs(self):
1040
        if self.config.fewshot_split is not None:
1041
1042
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1043
            return self.dataset[self.config.fewshot_split]
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
        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."
                )
1056
        else:
1057
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1058
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1059
                    f"[Task: {self.config.task}] "
1060
1061
1062
1063
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1064

KonradSzafer's avatar
KonradSzafer committed
1065
1066
1067
1068
1069
    @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
1070
        gen_prefix: Optional[str] = None,
1071
        question_suffix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1072
1073
1074
1075
1076
1077
1078
1079
    ) -> None:
        """Adds a target question to the labeled examples list.
        If fewshot_as_multiturn is True, or labeled_examples is empty, or the last entry is a system turn, appends the question as a new user entry.
        Otherwise, it is appended to the last user entry, ensuring that the conversation alternates between the user and the assistant.
        """
        if not fewshot_as_multiturn:
            # if no messages or last message is system, append as new user entry
            if len(labeled_examples) == 0 or labeled_examples[-1]["role"] == "system":
1080
                labeled_examples.append({"role": "user", "content": question + question_suffix} if question_suffix else {"role": "user", "content": question}  )
KonradSzafer's avatar
KonradSzafer committed
1081
1082
            # if last message is user, append to it to avoid two user messages in a row
            else:
1083
                labeled_examples[-1]["content"] += question + question_suffix if question_suffix else question
KonradSzafer's avatar
KonradSzafer committed
1084
1085
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1086
            labeled_examples.append({"role": "user", "content": question + question_suffix} if question_suffix else {"role": "user", "content": question}  )
Baber Abbasi's avatar
Baber Abbasi committed
1087
1088
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1089

lintangsutawika's avatar
lintangsutawika committed
1090
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1091
1092
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1093
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1094
1095
1096
1097
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1098
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1099
        gen_prefix: Optional[str] = None,
1100
        question_suffix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1101
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1102
1103
1104
1105
1106
1107
1108
        """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
1109
1110
1111
1112
1113
1114
        :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.
1115
1116
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1117
1118
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1119
1120
1121
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1122
1123
1124
1125
1126
1127
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1128
1129
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1130

KonradSzafer's avatar
KonradSzafer committed
1131
1132
1133
1134
1135
1136
1137
1138
1139
        # 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
1140
        else:
KonradSzafer's avatar
KonradSzafer committed
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
            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
1154
1155
1156
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1157
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1158
1159
1160
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1161
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1162
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1163
                )
lintangsutawika's avatar
lintangsutawika committed
1164
1165

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1166
1167
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1168
                # TODO: append prefill?
1169
1170
                if not labeled_examples:
                    return ""
1171
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1172
1173
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1174
1175
1176
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1177
                    gen_prefix=gen_prefix,
1178
                    question_suffix=question_suffix,
KonradSzafer's avatar
KonradSzafer committed
1179
1180
1181
1182
1183
1184
1185
                )
            # 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
1186
1187
1188
1189
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1190
                        gen_prefix=gen_prefix,
1191
                        question_suffix=question_suffix,
Baber Abbasi's avatar
Baber Abbasi committed
1192
1193
1194
1195
1196
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1197
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1198
1199
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1200
1201
1202
1203
1204
1205
                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
1206
1207
1208
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1209
                        gen_prefix=gen_prefix,
1210
                        question_suffix=question_suffix,
KonradSzafer's avatar
KonradSzafer committed
1211
1212
1213
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1214
1215
1216
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1217
                        gen_prefix=gen_prefix,
1218
                        question_suffix=question_suffix,
KonradSzafer's avatar
KonradSzafer committed
1219
1220
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1221
1222
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1223
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1224
            )
1225
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1226
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1227
1228
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1229
1230
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1231
1232
            if self.multiple_input:
                return labeled_examples
1233
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1234
                return labeled_examples + example + prefix
1235
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1236
                return [labeled_examples + ex + prefix for ex in example]
1237
1238
1239
            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
1240
                    return labeled_examples + choices[example] + prefix
1241
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1242
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1243

Baber Abbasi's avatar
Baber Abbasi committed
1244
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1245
        """Iterates over FilterEnsembles and applies them to instances"""
1246
1247
        if hasattr(self, "_filters"):
            for f in self._filters:
1248
                f.apply(self._instances)
1249
1250
1251
1252
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1253
    def should_decontaminate(self):
1254
        return self.config.should_decontaminate
1255

Baber Abbasi's avatar
Baber Abbasi committed
1256
    def doc_to_decontamination_query(self, doc: dict):
1257
        if self.config.should_decontaminate:
1258
1259
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1260
            else:
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
                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
                        )
                    )
1272

1273
    def _process_doc(self, doc: dict) -> dict:
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
        """
        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
1284
    def doc_to_text(self, doc, doc_to_text=None):
1285
1286
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1287
1288
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1289
        else:
1290
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1291

1292
        if isinstance(doc_to_text, int):
1293
            return doc_to_text
1294
        elif isinstance(doc_to_text, str):
1295
            if doc_to_text in self.features:
1296
                # if self.config.doc_to_choice is not None:
1297
1298
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1299
1300
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1301
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1302
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1303
1304
1305
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1306
        elif callable(doc_to_text):
1307
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1308
        # Used when applying a Promptsource template
1309
        elif hasattr(doc_to_text, "apply"):
1310
1311
1312
1313
1314
            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")
1315
                return self.config.fewshot_delimiter
1316
        else:
1317
            print(type(doc_to_text))
1318
            raise TypeError
1319

Yu Shi Jie's avatar
Yu Shi Jie committed
1320
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1321
1322
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1323
1324
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1325
        else:
1326
            doc_to_target = self.config.doc_to_target
1327

1328
        if isinstance(doc_to_target, int):
1329
            return doc_to_target
1330
        elif isinstance(doc_to_target, str):
1331
            if doc_to_target in self.features:
1332
                # if self.config.doc_to_choice is not None:
1333
1334
1335
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1336
            else:
lintangsutawika's avatar
lintangsutawika committed
1337
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1338
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1339
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1340
1341
1342
1343
1344
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1345
1346
1347
1348
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1349
1350
                else:
                    return target_string
1351
        elif isinstance(doc_to_target, list):
1352
            return doc_to_target
1353
        elif callable(doc_to_target):
1354
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1355
        # Used when applying a Promptsource template
1356
        elif hasattr(doc_to_target, "apply"):
1357
            applied_prompt = doc_to_target.apply(doc)
1358
1359
1360
1361
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1362
                return self.config.fewshot_delimiter
1363
1364
        else:
            raise TypeError
1365

Yu Shi Jie's avatar
Yu Shi Jie committed
1366
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1367
1368
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1369
1370
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1371
        elif self.config.doc_to_choice is None:
1372
1373
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1374
            doc_to_choice = self.config.doc_to_choice
1375

1376
        if isinstance(doc_to_choice, str):
1377
1378
1379
1380
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1381
        elif isinstance(doc_to_choice, list):
1382
            return doc_to_choice
1383
        elif isinstance(doc_to_choice, dict):
1384
1385
1386
1387
1388
1389
1390
            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
1391

1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
    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

1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
    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
1438
1439
1440
1441
1442
1443
1444
1445
    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
1446
1447
1448
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1449
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1450
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1451

1452
1453
        aux_arguments = None

1454
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1455
            arguments = (ctx, self.doc_to_target(doc))
1456
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1457
            arguments = (self.doc_to_target(doc),)
1458
        elif self.OUTPUT_TYPE == "multiple_choice":
1459
            choices = self.doc_to_choice(doc)
1460
            target_delimiter = self.config.target_delimiter
1461
1462
            if apply_chat_template:
                target_delimiter = ""
1463
1464
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1465
                # apply chat_template to choices if apply_chat_template
1466
                cont = self.doc_to_target(doc)
1467

1468
                arguments = [
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1479
                ]
1480
            else:
1481
                # Otherwise they are placed in the continuation
1482
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1483

1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # log(P(choice|ctx) / P(choice)) = log(P(choice|ctx)) - log(P(choice))
                # in other words normalizing by subtracting the unconditional logprob of each choice.
                aux_arguments = [("", f"{choice}") for choice in choices]

                arguments.extend(aux_arguments)

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

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

1508
1509
1510
1511
1512
1513
1514
1515
        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)},
            }

1516
1517
1518
1519
1520
1521
1522
        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":
1523
            request_list = [
1524
1525
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1526
                    doc=doc,
1527
                    arguments=arg,
1528
                    idx=i,
1529
1530
                    **kwargs,
                )
1531
                for i, arg in enumerate(arguments)
1532
            ]
1533
1534

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1535

lintangsutawika's avatar
lintangsutawika committed
1536
        return Instance(
1537
1538
1539
1540
1541
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1542
        )
1543
1544

    def process_results(self, doc, results):
1545
1546
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1547

1548
        result_dict = {}
1549
        use_metric = list(self._metric_fn_list.keys())
1550
1551
1552
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1553
1554
1555
1556
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1557
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1558
            (loglikelihood,) = results
1559
1560
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1561
            return {
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
                **(
                    {"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
1577
            }
1578
        elif self.OUTPUT_TYPE == "multiple_choice":
1579
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1580

1581
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1582
            choices = self.doc_to_choice(doc)
1583
1584
            completion_len = np.array([float(len(i)) for i in choices])

1585
1586
            if (
                2 * len(choices) == len(lls)
1587
                and "acc_mutual_info" in self._metric_fn_list.keys()
1588
1589
1590
1591
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
                lls_unconditional = lls[1::2]
1592
1593
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1594
1595
                # and this stores our "regular" conditional loglikelihoods
                lls = lls[::2]
1596

1597
1598
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1599

1600
1601
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1602
            else:
1603
                gold = self.doc_to_target(doc)
1604
1605

            gold_index_error = False
1606
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1607
1608
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1609
1610
                    gold_index_error = True
            else:
1611
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1612
                    gold = gold if gold < len(choices) else -100
1613
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1614
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1615

Lintang Sutawika's avatar
Lintang Sutawika committed
1616
                if gold == -100:
1617
1618
1619
1620
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1621
                    f"Label index was not in within range of available choices,"
1622
1623
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1624

1625
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1626
1627
                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
1628
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1629
1630
1631
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1632
                # 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
1633
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1634

Lintang Sutawika's avatar
Lintang Sutawika committed
1635
1636
1637
1638
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1639
            result_dict = {
1640
                **({"acc": acc} if "acc" in use_metric else {}),
1641
1642
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1643
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1644
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1645
1646
1647
1648
1649
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1650
1651
            }

1652
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1653
1654
1655
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1656
1657
1658
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1659
        elif self.OUTPUT_TYPE == "generate_until":
1660
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1661
            result = results[0]
1662
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1663
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1664
                # it assumes that doc_to_target returns a number.
1665
1666
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1667
1668
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1669
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1670
1671
1672
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
1673
            ):
Chris's avatar
Chris committed
1674
1675
                # cast gold to the same type as result
                gold = type(result)(gold)
1676

lintangsutawika's avatar
lintangsutawika committed
1677
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1678
1679
1680
1681
1682
                if self.multiple_target:
                    # in the case where we have multiple targets,
                    # return true if any are true
                    # TODO: this may break for multipLe_target, non zero-or-1 metrics
                    scores = []
haileyschoelkopf's avatar
haileyschoelkopf committed
1683
1684
1685
1686
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1687
1688
1689
1690
1691
1692
1693
1694
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1695
                    else:
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
1713
1714
1715
1716
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1717
                else:
1718
                    try:
1719
                        result_score = self._metric_fn_list[metric](
1720
1721
                            references=[gold],
                            predictions=[result],
1722
                            **self._metric_fn_kwargs[metric],
1723
                        )
1724
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1725
                        result_score = self._metric_fn_list[metric]([gold, result])
1726
1727
1728
1729
1730
1731
1732
                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
1733
        else:
lintangsutawika's avatar
lintangsutawika committed
1734
1735
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1736
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1737
            )
1738
1739
1740

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1741
    def aggregation(self) -> dict:
1742
1743
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1744
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1745
        return self._higher_is_better
1746

Baber Abbasi's avatar
Baber Abbasi committed
1747
1748
1749
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1750
1751
1752
1753
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1754
1755
1756
1757
1758
    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
1759
            f"num_samples={len(self.eval_docs)})"
1760
1761
        )

1762
1763

class MultipleChoiceTask(Task):
1764
    OUTPUT_TYPE = "loglikelihood"
1765

baberabb's avatar
baberabb committed
1766
    def doc_to_target(self, doc: dict) -> str:
1767
1768
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1769
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1770
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1771
1772
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1773
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1774
                doc=doc,
1775
                arguments=(ctx, " {}".format(choice)),
1776
                idx=i,
1777
1778
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1779
1780
            for i, choice in enumerate(doc["choices"])
        ]
1781

1782
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1783
1784
1785
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
        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
1797
    def higher_is_better(self) -> dict:
1798
1799
1800
1801
1802
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1803
    def aggregation(self) -> dict:
1804
1805
1806
1807
1808
1809
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1810
class PerplexityTask(Task):
1811
1812
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1813
    def has_training_docs(self) -> bool:
1814
1815
        return False

baberabb's avatar
baberabb committed
1816
    def fewshot_examples(self, k: int, rnd) -> List:
1817
1818
1819
1820
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1821
1822
        return []

baberabb's avatar
baberabb committed
1823
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1824
1825
1826
1827
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1828
1829
1830

        return ""

baberabb's avatar
baberabb committed
1831
    def higher_is_better(self) -> dict:
1832
1833
1834
1835
1836
1837
1838
1839
1840
        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
1841
    def doc_to_text(self, doc) -> str:
1842
1843
1844
1845
1846
        return ""

    def doc_to_target(self, doc):
        return doc

1847
1848
1849
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1850

lintangsutawika's avatar
lintangsutawika committed
1851
1852
1853
1854
1855
1856
1857
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1858

1859
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1860
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1861
1862
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1863
1864
1865
1866
1867
1868
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1869
    def aggregation(self) -> dict:
1870
1871
1872
1873
1874
1875
1876
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1877
    def count_bytes(cls, doc) -> int:
1878
1879
1880
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1881
    def count_words(cls, doc) -> int:
1882
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1883
        return len(re.split(r"\s+", doc))