task.py 72.3 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
    ) -> None:
405
        """Build a set of Instances for a task, and store them in task.instances"""
406
407
408
409

        # used with caching
        og_limit = limit

410
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
411
412
413
414
415
416
417
        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 ""
        )
418
        cache_key += f"-tokenizer{tokenizer_name}"
419

Baber Abbasi's avatar
Baber Abbasi committed
420
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
421
422
423
424
425
426
427
428
429
430
431
432
433

        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
434
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
435

436
        instances = []
437
438
439
440
441
442
443
444
445
446

        # 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(
447
448
449
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
450
451
452
453
454
455
456
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
457
        ):
458
            # sample fewshot context #TODO: need to offset doc_id by rank now!
459
            fewshot_ctx = self.fewshot_context(
460
                doc,
461
                0 if self.config.num_fewshot is None else self.config.num_fewshot,
KonradSzafer's avatar
KonradSzafer committed
462
463
464
                system_instruction,
                apply_chat_template,
                fewshot_as_multiturn,
465
                chat_template,
Baber Abbasi's avatar
Baber Abbasi committed
466
                gen_prefix=self.doc_to_prefix(doc),
467
            )
468

469
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
470
471
472
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
473
                metadata=(self.config["task"], doc_id, self.config.repeats),
474
                apply_chat_template=apply_chat_template,
475
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
476
            )
477
478
479
480

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

481
482
483
484
485
486
487
488
489
490
491
492
493
            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
494

495
496
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
497

498
499
500
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
    @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
517
            The number of times each instance in a dataset is inferred on. Defaults to 1,
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
            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

553
554
555
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
556
557
558
559
560
561
562
563
564
565
    @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))

566
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
567
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
568
569
570
571
572
573
574
        """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
575
576
577
578
579
        :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.
580
581
582
        :returns: str
            The fewshot context.
        """
583
        if rnd is None:
584
585
586
587
588
589
            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
590

591
        description = description if description else ""
592
593

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
594
            labeled_examples = ""
595
        else:
lintangsutawika's avatar
lintangsutawika committed
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
            # 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
620
            )
621
622

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
623
        return description + labeled_examples + example
624

625
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
626
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
627
628
        if hasattr(self, "_filters"):
            for f in self._filters:
629
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
630
631
632
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
633

baberabb's avatar
baberabb committed
634
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
635
        """Returns the config as a dictionary."""
636
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
637
        # (num_fewshot)
638
        return self.config.to_dict()
639

Baber Abbasi's avatar
Baber Abbasi committed
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
    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)

680
681
682
683
684
    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

685
686
687
688
689
690
691
    @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:
692
693
694
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
695
696

    def doc_iterator(
697
698
699
700
701
702
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
703
    ) -> Iterator[Tuple[int, Any]]:
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
        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),
            )
726
727
        return doc_iterator

728
729

class ConfigurableTask(Task):
730
    VERSION = "Yaml"
731
    OUTPUT_TYPE = None
732
    CONFIG = None
733
734

    def __init__(
735
736
737
738
739
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
740
    ) -> None:  # TODO no super() call here
741
        # Get pre-configured attributes
742
        self._config = self.CONFIG
743

744
        # Use new configurations if there was no preconfiguration
745
        if self.config is None:
746
            self._config = TaskConfig(**config)
747
748
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
749
            if config is not None:
750
                self._config.__dict__.update(config)
751

752
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
753
754
755
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
756

757
758
759
760
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

761
        if self.config.output_type is not None:
762
763
764
765
            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)}'"
                )
766
            self.OUTPUT_TYPE = self.config.output_type
767

768
769
770
771
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

772
773
774
775
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
776
777
778
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

779
780
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
781

782
783
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
784

785
786
787
788
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
789

790
        if self.config.metric_list is None:
791
            # TODO: handle this in TaskConfig.__post_init__ ?
792
793
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

794
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
795
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
796
                self._metric_fn_kwargs[metric_name] = {}
797
798
799
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
800
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
801
        else:
802
            for metric_config in self.config.metric_list:
803
804
805
806
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
807
808
809
810
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
811
812
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
813
                }
Chris's avatar
Chris committed
814
815
816
817
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
818

819
                if self.config.process_results is not None:
820
821
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
822
823
824
825
826
827
                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
828
829
830
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
831
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
832

833
                if "aggregation" in metric_config:
834
                    agg_name = metric_config["aggregation"]
835
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
836
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
837
                    elif callable(agg_name):  # noqa: E721
838
839
840
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
841
                else:
842
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
843
                    metric_agg = get_metric_aggregation(metric_name)
844
                    eval_logger.warning(
845
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
846
847
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
848
                    )
849
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
850

851
852
853
854
855
856
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
857
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
858
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
859
                        f"higher_is_better={is_higher_better(metric_name)}"
860
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
861
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
862

863
        self.download(self.config.dataset_kwargs)
864
865
866
        self._training_docs = None
        self._fewshot_docs = None

867
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
868
            self._filters = []
869
            for filter_config in self.config.filter_list:
870
871
872
873
874
875
876
877
878
                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
879
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
880
        else:
Baber Abbasi's avatar
Baber Abbasi committed
881
882
883
884
            # 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."
            )
885
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
886

887
888
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
889
            self.prompt = get_prompt(
890
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
891
            )
892
893
894
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
895
        if self.fewshot_docs() is not None:
896
897
898
899
            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
900
901
902
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
            )
            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)}"
                )
919

920
        self.task_docs = self.eval_docs
921

922
        # Test One Doc
923
        self.features = list(self.task_docs.features.keys())
924
925
        self.multiple_input = 0
        self.multiple_target = 0
926
        test_doc = self.task_docs[0]
927
        test_text = self.doc_to_text(test_doc)
928
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
929

930
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
931
            test_choice = self.doc_to_choice(test_doc)
932
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
933
                eval_logger.error("doc_to_choice must return list")
934
935
            else:
                num_choice = len(test_choice)
936

937
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
938
939
940
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
941
                self.multiple_input = num_choice
942
943
        else:
            test_choice = None
944

945
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
946
947
948
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
949
            self.multiple_target = len(test_target)
950
        else:
951
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
952
                test_target = test_choice[test_target]
953
            else:
lintangsutawika's avatar
lintangsutawika committed
954
                test_target = str(test_target)
955

956
957
958
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
959
            check_choices = [test_target]
960
961
962
963
        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 = (
964
965
                    True
                    if self.config.target_delimiter.rstrip()
966
                    != self.config.target_delimiter
967
                    else False
968
                )
969

970
                if delimiter_has_whitespace and choice_has_whitespace:
971
972
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
973
974
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
975
                    eval_logger.debug(
976
                        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'
977
978
                    )

Baber Abbasi's avatar
Baber Abbasi committed
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
    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 {},
            )
996

baberabb's avatar
baberabb committed
997
    def has_training_docs(self) -> bool:
998
        if self.config.training_split is not None:
999
1000
1001
1002
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1003
    def has_validation_docs(self) -> bool:
1004
        if self.config.validation_split is not None:
1005
1006
1007
1008
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1009
    def has_test_docs(self) -> bool:
1010
        if self.config.test_split is not None:
1011
1012
1013
1014
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1015
    def training_docs(self) -> datasets.Dataset:
1016
        if self.has_training_docs():
1017
1018
1019
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1020
                )
1021
            return self.dataset[self.config.training_split]
1022

baberabb's avatar
baberabb committed
1023
    def validation_docs(self) -> datasets.Dataset:
1024
        if self.has_validation_docs():
1025
1026
1027
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1028
                )
1029
            return self.dataset[self.config.validation_split]
1030

baberabb's avatar
baberabb committed
1031
    def test_docs(self) -> datasets.Dataset:
1032
        if self.has_test_docs():
1033
1034
1035
            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]
1036

1037
    def fewshot_docs(self):
1038
        if self.config.fewshot_split is not None:
1039
1040
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1041
            return self.dataset[self.config.fewshot_split]
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
        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."
                )
1054
        else:
1055
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1056
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1057
                    f"[Task: {self.config.task}] "
1058
1059
1060
1061
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1062

KonradSzafer's avatar
KonradSzafer committed
1063
1064
1065
1066
1067
    @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
1068
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1069
1070
1071
1072
1073
1074
1075
1076
    ) -> 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":
1077
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1078
1079
            # if last message is user, append to it to avoid two user messages in a row
            else:
1080
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1081
1082
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1083
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1084
1085
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1086

lintangsutawika's avatar
lintangsutawika committed
1087
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1088
1089
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1090
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1091
1092
1093
1094
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1095
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1096
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1097
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1098
1099
1100
1101
1102
1103
1104
        """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
1105
1106
1107
1108
1109
1110
        :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.
1111
1112
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1113
1114
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1115
1116
1117
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1118
1119
1120
1121
1122
1123
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1124
1125
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1126

KonradSzafer's avatar
KonradSzafer committed
1127
1128
1129
1130
1131
1132
1133
1134
1135
        # 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
1136
        else:
KonradSzafer's avatar
KonradSzafer committed
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
            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
1150
1151
1152
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1153
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1154
1155
1156
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1157
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1158
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1159
                )
lintangsutawika's avatar
lintangsutawika committed
1160
1161

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1162
1163
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1164
                # TODO: append prefill?
1165
1166
                if not labeled_examples:
                    return ""
1167
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1168
1169
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1170
1171
1172
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1173
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1174
1175
1176
1177
1178
1179
1180
                )
            # 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
1181
1182
1183
1184
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1185
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1186
1187
1188
1189
1190
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1191
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1192
1193
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1194
1195
1196
1197
1198
1199
                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
1200
1201
1202
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1203
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1204
1205
1206
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1207
1208
1209
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1210
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1211
1212
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1213
1214
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1215
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1216
            )
1217
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1218
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1219
1220
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1221
1222
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1223
1224
            if self.multiple_input:
                return labeled_examples
1225
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1226
                return labeled_examples + example + prefix
1227
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1228
                return [labeled_examples + ex + prefix for ex in example]
1229
1230
1231
            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
1232
                    return labeled_examples + choices[example] + prefix
1233
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1234
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1235

Baber Abbasi's avatar
Baber Abbasi committed
1236
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1237
        """Iterates over FilterEnsembles and applies them to instances"""
1238
1239
        if hasattr(self, "_filters"):
            for f in self._filters:
1240
                f.apply(self._instances)
1241
1242
1243
1244
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1245
    def should_decontaminate(self):
1246
        return self.config.should_decontaminate
1247

Baber Abbasi's avatar
Baber Abbasi committed
1248
    def doc_to_decontamination_query(self, doc: dict):
1249
        if self.config.should_decontaminate:
1250
1251
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1252
            else:
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
                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
                        )
                    )
1264

1265
    def _process_doc(self, doc: dict) -> dict:
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
        """
        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
1276
    def doc_to_text(self, doc, doc_to_text=None):
1277
1278
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1279
1280
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1281
        else:
1282
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1283

1284
        if isinstance(doc_to_text, int):
1285
            return doc_to_text
1286
        elif isinstance(doc_to_text, str):
1287
            if doc_to_text in self.features:
1288
                # if self.config.doc_to_choice is not None:
1289
1290
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1291
1292
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1293
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1294
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1295
1296
1297
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1298
        elif callable(doc_to_text):
1299
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1300
        # Used when applying a Promptsource template
1301
        elif hasattr(doc_to_text, "apply"):
1302
1303
1304
1305
1306
            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")
1307
                return self.config.fewshot_delimiter
1308
        else:
1309
            print(type(doc_to_text))
1310
            raise TypeError
1311

Yu Shi Jie's avatar
Yu Shi Jie committed
1312
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1313
1314
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1315
1316
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1317
        else:
1318
            doc_to_target = self.config.doc_to_target
1319

1320
        if isinstance(doc_to_target, int):
1321
            return doc_to_target
1322
        elif isinstance(doc_to_target, str):
1323
            if doc_to_target in self.features:
1324
                # if self.config.doc_to_choice is not None:
1325
1326
1327
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1328
            else:
lintangsutawika's avatar
lintangsutawika committed
1329
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1330
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1331
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1332
1333
1334
1335
1336
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1337
1338
1339
1340
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1341
1342
                else:
                    return target_string
1343
        elif isinstance(doc_to_target, list):
1344
            return doc_to_target
1345
        elif callable(doc_to_target):
1346
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1347
        # Used when applying a Promptsource template
1348
        elif hasattr(doc_to_target, "apply"):
1349
            applied_prompt = doc_to_target.apply(doc)
1350
1351
1352
1353
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1354
                return self.config.fewshot_delimiter
1355
1356
        else:
            raise TypeError
1357

Yu Shi Jie's avatar
Yu Shi Jie committed
1358
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1359
1360
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1361
1362
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1363
        elif self.config.doc_to_choice is None:
1364
1365
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1366
            doc_to_choice = self.config.doc_to_choice
1367

1368
        if isinstance(doc_to_choice, str):
1369
1370
1371
1372
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1373
        elif isinstance(doc_to_choice, list):
1374
            return doc_to_choice
1375
        elif isinstance(doc_to_choice, dict):
1376
1377
1378
1379
1380
1381
1382
            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
1383

1384
1385
1386
1387
1388
1389
1390
1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
    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

1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
    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
1430
1431
1432
1433
1434
1435
1436
1437
    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
1438
1439
1440
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1441
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1442
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1443

1444
1445
        aux_arguments = None

1446
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1447
            arguments = (ctx, self.doc_to_target(doc))
1448
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1449
            arguments = (self.doc_to_target(doc),)
1450
        elif self.OUTPUT_TYPE == "multiple_choice":
1451
            choices = self.doc_to_choice(doc)
1452
            target_delimiter = self.config.target_delimiter
1453
1454
            if apply_chat_template:
                target_delimiter = ""
1455
1456
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1457
                # apply chat_template to choices if apply_chat_template
1458
                cont = self.doc_to_target(doc)
1459

1460
                arguments = [
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1471
                ]
1472
            else:
1473
                # Otherwise they are placed in the continuation
1474
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1475

1476
1477
1478
1479
1480
1481
1482
1483
            # 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.
1484
1485
1486
1487
                # TODO: should these be strided? will have to modify the processing in process_results if so
                aux_arguments = [
                    ("", f"{target_delimiter}{choice}") for choice in choices
                ]
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502

                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)},
            }

1503
1504
1505
1506
1507
1508
1509
1510
        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)},
            }

1511
1512
1513
1514
1515
1516
1517
        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":
1518
            request_list = [
1519
1520
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1521
                    doc=doc,
1522
                    arguments=arg,
1523
                    idx=i,
1524
1525
                    **kwargs,
                )
1526
                for i, arg in enumerate(arguments)
1527
            ]
1528
1529

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1530

lintangsutawika's avatar
lintangsutawika committed
1531
        return Instance(
1532
1533
1534
1535
1536
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1537
        )
1538
1539

    def process_results(self, doc, results):
1540
1541
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1542

1543
        result_dict = {}
1544
        use_metric = list(self._metric_fn_list.keys())
1545
1546
1547
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1548
1549
1550
1551
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1552
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1553
            (loglikelihood,) = results
1554
1555
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1556
            return {
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
                **(
                    {"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
1572
            }
1573
        elif self.OUTPUT_TYPE == "multiple_choice":
1574
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1575

1576
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1577
            choices = self.doc_to_choice(doc)
1578
1579
            completion_len = np.array([float(len(i)) for i in choices])

1580
1581
            if (
                2 * len(choices) == len(lls)
1582
                and "acc_mutual_info" in self._metric_fn_list.keys()
1583
1584
1585
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1586
1587
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1588
1589
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1590
                # and this stores our "regular" conditional loglikelihoods
1591
                lls = lls[: len(choices)]
1592

1593
1594
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1595

1596
1597
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1598
            else:
1599
                gold = self.doc_to_target(doc)
1600
1601

            gold_index_error = False
1602
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1603
1604
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1605
1606
                    gold_index_error = True
            else:
1607
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1608
                    gold = gold if gold < len(choices) else -100
1609
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1610
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1611

Lintang Sutawika's avatar
Lintang Sutawika committed
1612
                if gold == -100:
1613
1614
1615
1616
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1617
                    f"Label index was not in within range of available choices,"
1618
1619
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1620

1621
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1622
1623
                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
1624
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1625
1626
1627
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1628
                # 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
1629
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1630

Lintang Sutawika's avatar
Lintang Sutawika committed
1631
1632
1633
1634
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1635
            result_dict = {
1636
                **({"acc": acc} if "acc" in use_metric else {}),
1637
1638
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1639
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1640
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1641
1642
1643
1644
1645
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1646
1647
            }

1648
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1649
1650
1651
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1652
1653
1654
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1655
        elif self.OUTPUT_TYPE == "generate_until":
1656
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1657
            result = results[0]
1658
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1659
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1660
                # it assumes that doc_to_target returns a number.
1661
1662
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1663
1664
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1665
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1666
1667
1668
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
1669
            ):
Chris's avatar
Chris committed
1670
1671
                # cast gold to the same type as result
                gold = type(result)(gold)
1672

lintangsutawika's avatar
lintangsutawika committed
1673
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1674
1675
1676
1677
1678
                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
1679
1680
1681
1682
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1683
1684
1685
1686
1687
1688
1689
1690
                    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
1691
                    else:
1692
1693
1694
1695
1696
1697
1698
1699
1700
1701
1702
1703
1704
1705
1706
1707
1708
1709
1710
1711
1712
                        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
1713
                else:
1714
                    try:
1715
                        result_score = self._metric_fn_list[metric](
1716
1717
                            references=[gold],
                            predictions=[result],
1718
                            **self._metric_fn_kwargs[metric],
1719
                        )
1720
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1721
                        result_score = self._metric_fn_list[metric]([gold, result])
1722
1723
1724
1725
1726
1727
1728
                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
1729
        else:
lintangsutawika's avatar
lintangsutawika committed
1730
1731
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1732
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1733
            )
1734
1735
1736

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1737
    def aggregation(self) -> dict:
1738
1739
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1740
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1741
        return self._higher_is_better
1742

Baber Abbasi's avatar
Baber Abbasi committed
1743
1744
1745
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1746
1747
1748
1749
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1750
1751
1752
1753
1754
    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
1755
            f"num_samples={len(self.eval_docs)})"
1756
1757
        )

1758
1759

class MultipleChoiceTask(Task):
1760
    OUTPUT_TYPE = "loglikelihood"
1761

baberabb's avatar
baberabb committed
1762
    def doc_to_target(self, doc: dict) -> str:
1763
1764
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1765
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1766
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1767
1768
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1769
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1770
                doc=doc,
1771
                arguments=(ctx, " {}".format(choice)),
1772
                idx=i,
1773
1774
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1775
1776
            for i, choice in enumerate(doc["choices"])
        ]
1777

1778
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1779
1780
1781
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1782
1783
1784
1785
1786
1787
1788
1789
1790
1791
1792
        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
1793
    def higher_is_better(self) -> dict:
1794
1795
1796
1797
1798
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1799
    def aggregation(self) -> dict:
1800
1801
1802
1803
1804
1805
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1806
class PerplexityTask(Task):
1807
1808
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1809
    def has_training_docs(self) -> bool:
1810
1811
        return False

baberabb's avatar
baberabb committed
1812
    def fewshot_examples(self, k: int, rnd) -> List:
1813
1814
1815
1816
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1817
1818
        return []

baberabb's avatar
baberabb committed
1819
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1820
1821
1822
1823
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1824
1825
1826

        return ""

baberabb's avatar
baberabb committed
1827
    def higher_is_better(self) -> dict:
1828
1829
1830
1831
1832
1833
1834
1835
1836
        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
1837
    def doc_to_text(self, doc) -> str:
1838
1839
1840
1841
1842
        return ""

    def doc_to_target(self, doc):
        return doc

1843
1844
1845
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1846

lintangsutawika's avatar
lintangsutawika committed
1847
1848
1849
1850
1851
1852
1853
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1854

1855
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1856
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1857
1858
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1859
1860
1861
1862
1863
1864
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1865
    def aggregation(self) -> dict:
1866
1867
1868
1869
1870
1871
1872
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1873
    def count_bytes(cls, doc) -> int:
1874
1875
1876
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1877
    def count_words(cls, doc) -> int:
1878
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1879
        return len(re.split(r"\s+", doc))