task.py 58.1 KB
Newer Older
Baber's avatar
cleanup  
Baber committed
1
2
from __future__ import annotations

3
import abc
4
import ast
lintangsutawika's avatar
lintangsutawika committed
5
import logging
6
import random
7
8
import re
from collections.abc import Callable
9
from copy import deepcopy
10
from typing import (
Baber's avatar
cleanup  
Baber committed
11
    TYPE_CHECKING,
12
13
14
    Any,
    Literal,
)
15
16
17

import datasets
import numpy as np
18
from tqdm import tqdm
Baber's avatar
Baber committed
19
from typing_extensions import deprecated
20
21

from lm_eval import utils
22
23
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
24
from lm_eval.api.utils import check_gold_index_error
25
from lm_eval.caching.cache import load_from_cache, save_to_cache
Baber's avatar
Baber committed
26
27
from lm_eval.config.metric import MetricConfig
from lm_eval.config.task import TaskConfig
28
29
from lm_eval.filters import build_filter_ensemble

30

31
32
33
34
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
35
    "generate_until",
36
37
]

Baber's avatar
cleanup  
Baber committed
38
if TYPE_CHECKING:
Baber's avatar
Baber committed
39
    pass
Baber's avatar
cleanup  
Baber committed
40
41


Lintang Sutawika's avatar
Lintang Sutawika committed
42
eval_logger = logging.getLogger(__name__)
43

lintangsutawika's avatar
lintangsutawika committed
44

45
46
47
48
49
50
51
52
53
54
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)
    """

Baber's avatar
cleanup  
Baber committed
55
    VERSION: int | str | None = None
56

57
58
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
Baber's avatar
cleanup  
Baber committed
59
    DATASET_PATH: str | None = None
60
61

    # The name of a subset within `DATASET_PATH`.
Baber's avatar
cleanup  
Baber committed
62
    DATASET_NAME: str | None = None
63

Baber's avatar
cleanup  
Baber committed
64
    OUTPUT_TYPE: OutputType | None = None
lintangsutawika's avatar
lintangsutawika committed
65

66
67
    def __init__(
        self,
Baber's avatar
cleanup  
Baber committed
68
69
70
71
        data_dir: str | None = None,
        cache_dir: str | None = None,
        download_mode: datasets.DownloadMode | None = None,
        config: Mapping | None = None,  # Union[dict, TaskConfig]
Ethan Smith's avatar
Ethan Smith committed
72
    ) -> None:
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
        """
        :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)
Baber's avatar
cleanup  
Baber committed
95
96
97
        self._training_docs: list | None = None
        self._fewshot_docs: list | None = None
        self._instances: list[Instance] | None = None
98

99
        self._config: TaskConfig = TaskConfig.from_yaml({**config})
100

101
        self._filters = [build_filter_ensemble("none", [("take_first", None)])]
Baber's avatar
cleanup  
Baber committed
102
        self.fewshot_rnd: random.Random | None = (
103
104
            None  # purposely induce errors in case of improper usage
        )
105

106
107
    def download(
        self,
Baber's avatar
cleanup  
Baber committed
108
109
        data_dir: str | None = None,
        cache_dir: str | None = None,
110
111
        download_mode=None,
    ) -> None:
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
        """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.
        """
136
137
138
139
140
141
142
        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,
        )
143

144
    @property
145
    def config(self) -> TaskConfig:
146
147
148
        """Returns the TaskConfig associated with this class."""
        return self._config

149
    @abc.abstractmethod
Baber's avatar
Baber committed
150
    def has_training_docs(self) -> bool:
151
152
153
154
        """Whether the task has a training set"""
        pass

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

    @abc.abstractmethod
Baber's avatar
Baber committed
160
    def has_test_docs(self) -> bool:
161
162
163
        """Whether the task has a test set"""
        pass

164
    def training_docs(self) -> Iterable:
165
166
167
168
169
170
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

171
    def validation_docs(self) -> Iterable:
172
173
174
175
176
177
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

178
    def test_docs(self) -> Iterable:
179
180
181
182
183
184
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

185
    def fewshot_docs(self) -> Iterable:
186
187
188
189
190
191
192
193
194
        """
        :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
195
196
197
198
199
            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."
                )
200
201
            return self.test_docs()

202
    def _process_doc(self, doc: dict) -> dict:
203
204
205
206
207
208
209
210
211
        """
        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
212

213
    @property
Baber's avatar
cleanup  
Baber committed
214
    def instances(self) -> list[Instance]:
215
216
217
218
219
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

Baber's avatar
Baber committed
220
    def fewshot_examples(self, k, rnd) -> Iterable[dict]:
221
222
223
224
225
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

Baber's avatar
cleanup  
Baber committed
226
    def doc_to_decontamination_query(self, doc: dict):
227
        raise NotImplementedError(
228
229
230
231
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
232
    def doc_to_text(self, doc: dict) -> str:
233
234
235
        pass

    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
236
    def doc_to_target(self, doc: dict) -> str | int:
237
238
        pass

239
    # not an abstractmethod because not every language-only task has to implement this
Baber's avatar
cleanup  
Baber committed
240
    def doc_to_image(self, doc: dict):
241
242
        raise NotImplementedError

Baber's avatar
cleanup  
Baber committed
243
    def doc_to_audio(self, doc: dict):
244
245
        raise NotImplementedError

Baber's avatar
cleanup  
Baber committed
246
    def doc_to_prefix(self, doc: dict) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
247
248
        return ""

249
250
    def build_all_requests(
        self,
251
        *,
Baber's avatar
cleanup  
Baber committed
252
253
        limit: int | None = None,
        samples: list[int] | None = None,
254
255
256
257
        rank: int = 0,
        world_size: int = 1,
        cache_requests: bool = False,
        rewrite_requests_cache: bool = False,
Baber's avatar
cleanup  
Baber committed
258
        system_instruction: str | None = None,
259
260
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
261
        chat_template: Callable | None = None,
262
        tokenizer_name: str = "",
263
    ) -> None:
264
        """Build a set of Instances for a task, and store them in task.instances"""
265
266
267
268

        # used with caching
        og_limit = limit

269
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
270
271
272
273
274
275
276
        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 ""
        )
277
        cache_key += f"-tokenizer{tokenizer_name}"
278

Baber Abbasi's avatar
Baber Abbasi committed
279
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
280
281
282
283
284
285
286
287
288
289
290
291
292

        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
293
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
294

295
        instances = []
296
297
298
299
300
301
302
303
304
305

        # 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(
306
307
308
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
309
310
311
312
313
314
315
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
316
        ):
317
            # sample fewshot context #TODO: need to offset doc_id by rank now!
318
            fewshot_ctx = self.fewshot_context(
319
                doc,
320
321
322
323
324
325
326
                num_fewshot=0
                if self.config.num_fewshot is None
                else self.config.num_fewshot,
                system_instruction=system_instruction,
                apply_chat_template=apply_chat_template,
                fewshot_as_multiturn=fewshot_as_multiturn,
                chat_template=chat_template,
Baber Abbasi's avatar
Baber Abbasi committed
327
                gen_prefix=self.doc_to_prefix(doc),
328
            )
329

330
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
331
332
333
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
334
                metadata=(self.config["task"], doc_id, self.config.repeats),
335
                apply_chat_template=apply_chat_template,
336
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
337
            )
338
339
340
341

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

342
343
344
345
346
347
348
349
350
351
352
353
354
            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
355

356
357
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
358

359
360
361
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

362
    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
363
    def construct_requests(self, doc: dict, ctx: list[dict] | str, **kwargs):
364
365
366
367
368
369
370
371
372
373
374
375
376
377
        """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
378
            The number of times each instance in a dataset is inferred on. Defaults to 1,
379
380
381
382
383
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
384
    def process_results(self, doc: dict, results: list) -> dict[str, Any]:
385
386
387
388
389
390
391
392
393
        """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.
        """
394
        raise NotImplementedError
395

Baber's avatar
Baber committed
396
    @deprecated("not used anymore")
397
398
399
400
401
402
    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
        """
Baber's avatar
cleanup  
Baber committed
403
        return True
404

Baber's avatar
Baber committed
405
    @deprecated("not used anymore")
406
407
408
409
410
411
    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
        """
Baber's avatar
cleanup  
Baber committed
412
        return True
413

414
415
416
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
417
    @classmethod
Baber's avatar
Baber committed
418
    def count_bytes(cls, doc: str) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
419
420
421
422
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
423
    def count_words(cls, doc: str) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
424
425
426
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

427
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
428
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
429
430
431
432
433
434
435
        """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
436
437
438
439
440
        :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.
441
442
443
        :returns: str
            The fewshot context.
        """
444
        if rnd is None:
445
446
447
448
449
450
            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
451

452
        description = description if description else ""
453
454

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
455
            labeled_examples = ""
456
        else:
lintangsutawika's avatar
lintangsutawika committed
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
            # 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
481
            )
482
483

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
484
        return description + labeled_examples + example
485

Baber's avatar
cleanup  
Baber committed
486
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
487
        """Iterates over FilterEnsembles and applies them to instances"""
Baber's avatar
cleanup  
Baber committed
488
        if hasattr(self, "_filters") and self._instances:
lintangsutawika's avatar
lintangsutawika committed
489
            for f in self._filters:
490
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
491
        else:
Baber's avatar
cleanup  
Baber committed
492
493
494
            eval_logger.warning(
                "No filter defined or no instances, passing through instances"
            )
lintangsutawika's avatar
lintangsutawika committed
495
            return self._instances
496

baberabb's avatar
baberabb committed
497
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
498
        """Returns the config as a dictionary."""
499
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
500
        # (num_fewshot)
501
        return self.config.to_dict()
502

Baber Abbasi's avatar
Baber Abbasi committed
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if update:
            current_value = getattr(self._config, key, {})
            if not isinstance(current_value, dict):
                raise TypeError(
                    f"Expected a dict for key '{key}', got {type(current_value).__name__} instead."
                )
            current_value.update(value)
        else:
            setattr(self._config, key, value)

    def override_metric(self, metric_name: str) -> None:
        """
        Override the default metrics used for evaluation with custom metrics.

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
Baber's avatar
Baber committed
522
523
524
525
526
527
528
        # if not isinstance(self, ConfigurableTask):
        #     self.process_results = lambda x, y: {metric_name: get_metric(metric_name)}
        #     self.aggregation = lambda: {
        #         metric_name: get_metric_aggregation(metric_name)
        #     }
        setattr(self._config, "metric_list", [MetricConfig(name=metric_name)])
        setattr(self._config, "process_results", lambda *args: {"bypass": 0})
Baber Abbasi's avatar
Baber Abbasi committed
529

Baber's avatar
cleanup  
Baber committed
530
    def set_fewshot_seed(self, seed: int | None = None) -> None:
531
532
533
534
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

535
    @property
Baber's avatar
cleanup  
Baber committed
536
    def eval_docs(self) -> datasets.Dataset | Iterable[dict]:
537
538
539
540
541
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
542
543
544
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
545
546

    def doc_iterator(
547
548
549
        self,
        *,
        rank: int = 0,
Baber's avatar
cleanup  
Baber committed
550
        limit: int | None = None,
551
        world_size: int = 1,
Baber's avatar
cleanup  
Baber committed
552
553
        samples: list[int] | None = None,
    ) -> Iterator[tuple[int, Any]]:
554
555
        if samples:
            n = len(self.eval_docs)
Baber's avatar
cleanup  
Baber committed
556
            assert all(e < n for e in samples), (
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
                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),
            )
576
577
        return doc_iterator

578
579

class ConfigurableTask(Task):
580
    VERSION = "Yaml"
581
    OUTPUT_TYPE = None
582
    CONFIG = None
583
584

    def __init__(
585
586
587
588
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
Baber's avatar
cleanup  
Baber committed
589
        config: dict | None = None,
Baber's avatar
Baber committed
590
    ) -> None:
591
        # Get pre-configured attributes
592
        self._config = self.CONFIG
593

594
        # Use new configurations if there was no preconfiguration
595
        if self.config is None:
596
            self._config = TaskConfig(**config)
597
598
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
599
            if config is not None:
600
                self._config.__dict__.update(config)
601

602
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
603
604
605
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
606

Baber's avatar
cleanup  
Baber committed
607
608
        if isinstance(self.config.metadata, dict) and "version" in self.config.metadata:
            self.VERSION = self.config.metadata["version"]
609

610
        if self.config.output_type is not None:
611
612
613
614
            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)}'"
                )
615
            self.OUTPUT_TYPE = self.config.output_type
616

617
618
619
620
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

621
622
623
624
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
625
626
627
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

628
629
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
630

631
632
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
633

634
        # self.metric_list: list[MetricConfig] = self.config.get_metrics
635

636
        self.download(self.config.dataset_kwargs)
637
638
639
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
640
        self._filters = self.config.get_filters
Baber's avatar
Baber committed
641

Baber's avatar
Baber committed
642
643
644
645
646
647
648
        # if self.config.use_prompt is not None:
        #     eval_logger.info(f"loading prompt {self.config.use_prompt}")
        #     self.prompt = get_prompt(
        #         self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
        #     )
        # else:
        #     self.prompt = None
649

650
651
652
653
        if (
            self.config.fewshot_cfg.num_fewshot() > 0
            and self.fewshot_docs() is not None
        ):
Baber's avatar
Baber committed
654
655
656
            self.fewshot_rnd = random.Random()
            self.sampler = self.config.fewshot_cfg.init_sampler(
                list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
657
            )
658
        self.task_docs = self.eval_docs
659

660
        # Test One Doc
Baber's avatar
Baber committed
661
        self.features: list[str] = list(self.task_docs.features.keys())
Baber's avatar
Baber committed
662
        self.multiple_input = self.config.multiple_input
663
        self.multiple_target = 0
664
        test_doc = self.task_docs[0]
665
        test_text = self.doc_to_text(test_doc)
666
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
667

668
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
669
            test_choice = self.doc_to_choice(test_doc)
670
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
671
                eval_logger.error("doc_to_choice must return list")
672
673
            else:
                num_choice = len(test_choice)
674

675
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
676
677
678
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
679
                self.multiple_input = num_choice
680
681
        else:
            test_choice = None
682

683
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
684
685
686
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
687
            self.multiple_target = len(test_target)
688
        else:
689
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
690
                test_target = test_choice[test_target]
691
            else:
lintangsutawika's avatar
lintangsutawika committed
692
                test_target = str(test_target)
693

Baber's avatar
cleanup  
Baber committed
694
        check_choices = test_choice if test_choice is not None else [test_target]
695
696
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
Baber's avatar
cleanup  
Baber committed
697
                choice_has_whitespace = choice[0].isspace()
698
                delimiter_has_whitespace = (
Baber's avatar
cleanup  
Baber committed
699
                    self.config.target_delimiter.rstrip()
700
                    != self.config.target_delimiter
701
                )
702

703
                if delimiter_has_whitespace and choice_has_whitespace:
704
705
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
706
707
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
708
                    eval_logger.debug(
709
                        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'
710
711
                    )

Baber Abbasi's avatar
Baber Abbasi committed
712
    def download(
Baber's avatar
cleanup  
Baber committed
713
        self, dataset_kwargs:dict[str, Any] | None = None, **kwargs
Baber Abbasi's avatar
Baber Abbasi committed
714
    ) -> None:
Baber Abbasi's avatar
Baber Abbasi committed
715
716
        from packaging.version import parse as vparse

717
718
719
720
        self.config.dataset_kwargs, self.config.metadata = (
            self.config.dataset_kwargs or {},
            self.config.metadata or {},
        )
Baber Abbasi's avatar
Baber Abbasi committed
721
722
        if dataset_kwargs and vparse(datasets.__version__) >= vparse("4.0.0"):
            dataset_kwargs.pop("trust_remote_code", None)
723
        if isinstance(df := self.config.custom_dataset, Callable):
Baber Abbasi's avatar
Baber Abbasi committed
724
725
726
727
            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."
            )
728
            self.dataset = df(**(self.config.dataset_kwargs | self.config.metadata))
Baber Abbasi's avatar
Baber Abbasi committed
729
730
        else:
            self.dataset = datasets.load_dataset(
731
732
733
                path=self.config.dataset_path,
                name=self.config.dataset_name,
                **self.config.dataset_kwargs,
Baber Abbasi's avatar
Baber Abbasi committed
734
            )
735

baberabb's avatar
baberabb committed
736
    def has_training_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
737
        return self.config.training_split is not None
738

baberabb's avatar
baberabb committed
739
    def has_validation_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
740
        return self.config.validation_split is not None
741

baberabb's avatar
baberabb committed
742
    def has_test_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
743
        return self.config.test_split is not None
744

Baber's avatar
cleanup  
Baber committed
745
    def training_docs(self) -> datasets.Dataset | None:
746
        if self.has_training_docs():
747
748
749
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
750
                )
751
            return self.dataset[self.config.training_split]
752

Baber's avatar
cleanup  
Baber committed
753
    def validation_docs(self) -> datasets.Dataset | None:
754
        if self.has_validation_docs():
755
756
757
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
758
                )
759
            return self.dataset[self.config.validation_split]
760

Baber's avatar
cleanup  
Baber committed
761
    def test_docs(self) -> datasets.Dataset | None:
762
        if self.has_test_docs():
763
764
765
            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]
766

767
    def fewshot_docs(self):
Baber's avatar
Baber committed
768
769
770
771
772
773
        docs = self.config.fewshot_cfg.get_docs(self.dataset)

        if docs is not None:
            return docs

        # Fallback to parent implementation
Baber's avatar
cleanup  
Baber committed
774
775
776
777
778
779
780
781
782
783
        if (
            (_num_fewshot := self.config.num_fewshot)
            and isinstance(_num_fewshot, int)
            and _num_fewshot > 0
        ):
            eval_logger.warning(
                f"[Task: {self.config.task}] "
                "num_fewshot > 0 but no fewshot source configured. "
                "Using preconfigured rule."
            )
Baber's avatar
Baber committed
784
785

        return super().fewshot_docs()
786

KonradSzafer's avatar
KonradSzafer committed
787
788
    @staticmethod
    def append_target_question(
Baber's avatar
cleanup  
Baber committed
789
        labeled_examples: list[dict[str, str]],
KonradSzafer's avatar
KonradSzafer committed
790
791
        question: str,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
792
        gen_prefix: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
793
794
795
796
797
798
799
800
    ) -> 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":
801
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
802
803
            # if last message is user, append to it to avoid two user messages in a row
            else:
804
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
805
806
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
807
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
808
809
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
810

lintangsutawika's avatar
lintangsutawika committed
811
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
812
813
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
814
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
815
        num_fewshot: int,
Baber's avatar
cleanup  
Baber committed
816
        system_instruction: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
817
818
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
819
820
821
        chat_template: Callable | None = None,
        gen_prefix: str | None = None,
    ) -> str | list[str] | None:
lintangsutawika's avatar
lintangsutawika committed
822
823
824
825
826
827
828
        """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
829
830
831
832
833
834
        :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.
835
836
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
837
838
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
839
840
841
        :returns: str
            The fewshot context.
        """
Baber's avatar
cleanup  
Baber committed
842
        labeled_examples = [] if apply_chat_template else ""
KonradSzafer's avatar
KonradSzafer committed
843
844

        # get task description
845
846
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
847

KonradSzafer's avatar
KonradSzafer committed
848
849
850
851
852
853
854
855
856
        # 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
857
        else:
KonradSzafer's avatar
KonradSzafer committed
858
859
860
861
862
863
864
865
866
867
868
869
870
            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
871
872
873
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
874
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
875
876
877
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
878
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
879
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
880
                )
lintangsutawika's avatar
lintangsutawika committed
881
882

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
883
884
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
885
                # TODO: append prefill?
886
887
                if not labeled_examples:
                    return ""
888
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
889
890
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
891
892
893
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
894
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
895
896
897
898
899
900
901
                )
            # 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
902
903
904
905
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
906
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
907
908
909
910
911
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber's avatar
cleanup  
Baber committed
912
                            add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
913
914
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
915
916
917
918
919
920
                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
921
922
923
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
924
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
925
926
927
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
928
929
930
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
931
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
932
933
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
934
935
            return chat_template(
                labeled_examples,
Baber's avatar
cleanup  
Baber committed
936
                add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
937
            )
938
        else:
Baber Abbasi's avatar
Baber Abbasi committed
939
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
940
941
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
942
943
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
944
945
            if self.multiple_input:
                return labeled_examples
946
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
947
                return labeled_examples + example + prefix
948
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
949
                return [labeled_examples + ex + prefix for ex in example]
950
951
952
            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
953
                    return labeled_examples + choices[example] + prefix
954
                else:
Baber Abbasi's avatar
Baber Abbasi committed
955
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
956

Baber's avatar
cleanup  
Baber committed
957
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
958
        """Iterates over FilterEnsembles and applies them to instances"""
959
        if hasattr(self, "_filters") and self._instances:
960
            for f in self._filters:
961
                f.ensemble.apply(self._instances)
962
        else:
963
964
965
            eval_logger.warning(
                "No filter defined or instances found. Passing through instances"
            )
966
967
            return self._instances

968
    def should_decontaminate(self):
969
        return self.config.should_decontaminate
970

Baber Abbasi's avatar
Baber Abbasi committed
971
    def doc_to_decontamination_query(self, doc: dict):
972
        if self.config.should_decontaminate:
973
974
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
975
            else:
976
977
978
979
980
981
982
983
984
985
986
                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
                        )
                    )
987

988
    def _process_doc(self, doc: dict) -> dict:
989
990
991
992
993
994
995
996
997
998
        """
        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

999
1000
1001
    def doc_to_text(
        self, doc: dict, doc_to_text: int | str | Callable[..., str] | None = None
    ) -> str:
Baber's avatar
Baber committed
1002
1003
        # if self.prompt is not None:
        #     doc_to_text = self.prompt
1004
        doc_to_text = doc_to_text or self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1005

1006
        if isinstance(doc_to_text, int):
1007
            return doc_to_text
1008
        elif isinstance(doc_to_text, str):
1009
            if doc_to_text in self.features:
1010
                # if self.config.doc_to_choice is not None:
1011
1012
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1013
1014
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1015
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
Baber committed
1016
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1017
1018
1019
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1020
        elif callable(doc_to_text):
1021
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1022
        # Used when applying a Promptsource template
Baber's avatar
Baber committed
1023
1024
1025
1026
1027
1028
1029
        # elif hasattr(doc_to_text, "apply"):
        #     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")
        #         return self.config.fewshot_delimiter
1030
        else:
1031
            print(type(doc_to_text))
1032
            raise TypeError
1033

Baber's avatar
cleanup  
Baber committed
1034
    def doc_to_target(self, doc: dict, doc_to_target=None) -> int | str | list[int]:
Baber's avatar
Baber committed
1035
1036
1037
        # if self.prompt is not None:
        #     doc_to_target = self.prompt
        if doc_to_target is not None:
Yu Shi Jie's avatar
Yu Shi Jie committed
1038
            doc_to_target = doc_to_target
1039
        else:
1040
            doc_to_target = self.config.doc_to_target
1041

1042
        if isinstance(doc_to_target, int):
1043
            return doc_to_target
1044
        elif isinstance(doc_to_target, str):
1045
            if doc_to_target in self.features:
1046
                # if self.config.doc_to_choice is not None:
1047
1048
1049
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1050
            else:
lintangsutawika's avatar
lintangsutawika committed
1051
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
Baber committed
1052
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1053
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1054
1055
1056
1057
1058
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1059
1060
1061
1062
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1063
1064
                else:
                    return target_string
1065
        elif isinstance(doc_to_target, list):
1066
            return doc_to_target
1067
        elif callable(doc_to_target):
1068
            return doc_to_target(doc)
Baber's avatar
Baber committed
1069
1070
1071
1072
1073
1074
1075
1076
        # # Used when applying a Promptsource template
        # elif hasattr(doc_to_target, "apply"):
        #     applied_prompt = doc_to_target.apply(doc)
        #     if len(applied_prompt) == 2:
        #         return applied_prompt[1]
        #     else:
        #         eval_logger.warning("Applied prompt returns empty string")
        #         return self.config.fewshot_delimiter
1077
1078
        else:
            raise TypeError
1079

Baber's avatar
cleanup  
Baber committed
1080
    def doc_to_choice(
Baber's avatar
Baber committed
1081
1082
        self,
        doc: dict,
Baber's avatar
cleanup  
Baber committed
1083
1084
        doc_to_choice: str | list | dict | Callable[..., list[str]] | None = None,
    ) -> list[str]:
Baber's avatar
Baber committed
1085
1086
1087
        # if self.prompt is not None:
        #     doc_to_choice = self.prompt
        if doc_to_choice is not None:
Yu Shi Jie's avatar
Yu Shi Jie committed
1088
            doc_to_choice = doc_to_choice
1089
        elif self.config.doc_to_choice is None:
1090
            eval_logger.error("doc_to_choice was called but not set in config")
Baber's avatar
Baber committed
1091
            doc_to_choice = None
1092
        else:
1093
            doc_to_choice = self.config.doc_to_choice
1094

1095
        if isinstance(doc_to_choice, str):
1096
1097
1098
1099
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1100
        elif isinstance(doc_to_choice, list):
1101
            return doc_to_choice
1102
        elif isinstance(doc_to_choice, dict):
1103
1104
1105
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
Baber's avatar
Baber committed
1106
1107
        # elif hasattr(doc_to_choice, "get_answer_choices_list"):
        #     return doc_to_choice.get_answer_choices_list(doc)
1108
1109
        else:
            raise TypeError
1110

Baber's avatar
cleanup  
Baber committed
1111
    def doc_to_image(self, doc: dict, doc_to_image=None) -> int | str | list | None:
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

Baber's avatar
cleanup  
Baber committed
1134
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> int | str | list | None:
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber's avatar
cleanup  
Baber committed
1157
    def doc_to_prefix(self, doc: dict) -> str | None:
Baber Abbasi's avatar
Baber Abbasi committed
1158
1159
1160
1161
1162
1163
1164
        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
1165
1166
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
Baber's avatar
cleanup  
Baber committed
1167
    ) -> list[Instance] | Instance:
1168
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1169
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1170

1171
1172
        aux_arguments = None

1173
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1174
            arguments = (ctx, self.doc_to_target(doc))
1175
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1176
            arguments = (self.doc_to_target(doc),)
1177
        elif self.OUTPUT_TYPE == "multiple_choice":
1178
            choices = self.doc_to_choice(doc)
1179
            target_delimiter = self.config.target_delimiter
1180
1181
            if apply_chat_template:
                target_delimiter = ""
1182
1183
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1184
                # apply chat_template to choices if apply_chat_template
1185
                cont = self.doc_to_target(doc)
1186

1187
                arguments = [
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1198
                ]
1199
            else:
1200
                # Otherwise they are placed in the continuation
1201
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1202

1203
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1204
            if "acc_mutual_info" in [m.metric_name for m in self.config._metric_list]:
1205
1206
1207
1208
1209
1210
                # 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.
1211
1212
1213
1214
                # 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
                ]
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229

                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)},
            }

1230
1231
1232
1233
1234
1235
1236
1237
        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)},
            }

1238
1239
1240
1241
1242
1243
1244
        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":
1245
            request_list = [
1246
1247
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1248
                    doc=doc,
1249
                    arguments=arg,
1250
                    idx=i,
1251
1252
                    **kwargs,
                )
1253
                for i, arg in enumerate(arguments)
1254
            ]
1255
1256

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1257

lintangsutawika's avatar
lintangsutawika committed
1258
        return Instance(
1259
1260
1261
1262
1263
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1264
        )
1265

1266
    def process_results(self, doc: dict, results: list) -> dict[str, Any]:
1267
1268
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1269

1270
        result_dict = {}
1271
        use_metric = list(m.metric_name for m in self.config._metric_list)
1272
1273
1274
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1275
1276
1277
1278
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1279
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
1280
1281
1282
1283
1284
1285
            (loglikelihood, *_) = results
            assert isinstance(_target := self.doc_to_target(doc), str), (
                "Require target to be a string for loglikelihood_rolling"
            )
            _words = self.count_words(_target)
            _bytes = self.count_bytes(_target)
haileyschoelkopf's avatar
haileyschoelkopf committed
1286
            return {
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
                **(
                    {"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
1302
            }
1303
        elif self.OUTPUT_TYPE == "multiple_choice":
1304
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1305

Baber's avatar
cleanup  
Baber committed
1306
            # retrieve choices in list[str] form, to compute choice lengths, etc.
1307
            choices = self.doc_to_choice(doc)
1308
1309
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1310
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1311
1312
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1313
1314
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1315
1316
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1317
                # and this stores our "regular" conditional loglikelihoods
1318
                lls = lls[: len(choices)]
Baber's avatar
Baber committed
1319
1320
            else:
                lls_unconditional = None
1321

1322
1323
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1324

1325
1326
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1327
            else:
1328
                gold = self.doc_to_target(doc)
1329

1330
            gold, gold_index_error = check_gold_index_error(choices, gold)
1331
1332
1333

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1334
                    f"Label index was not in within range of available choices,"
1335
1336
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1337

1338
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1339
1340
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
Baber's avatar
cleanup  
Baber committed
1341
                exact_match = int(any(is_greedy[i] if i != -100 else 0 for i in gold))
lintangsutawika's avatar
lintangsutawika committed
1342
1343
1344
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1345
                # 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
1346
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1347

Lintang Sutawika's avatar
Lintang Sutawika committed
1348
1349
1350
1351
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1352
            result_dict = {
1353
                **({"acc": acc} if "acc" in use_metric else {}),
1354
1355
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1356
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1357
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1358
1359
1360
1361
1362
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1363
1364
            }

1365
            if "acc_mutual_info" in use_metric:
Baber's avatar
Baber committed
1366
1367
1368
                assert lls_unconditional is not None, (
                    "lls_unconditional should not be None if acc_mutual_info is in use_metric"
                )
lintangsutawika's avatar
lintangsutawika committed
1369
1370
1371
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1372
1373
1374
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1375
        elif self.OUTPUT_TYPE == "generate_until":
1376
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1377
            result = results[0]
Baber's avatar
cleanup  
Baber committed
1378
            for metric in self._metric_fn_list:
1379
1380
1381
1382
1383
1384
1385
1386
                try:
                    result_score = self._metric_fn_list[metric](
                        references=[gold] if not isinstance(gold, list) else gold,
                        predictions=[result],
                        **self._metric_fn_kwargs[metric],
                    )
                except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
                    result_score = self._metric_fn_list[metric]([gold, result])
1387
1388
1389
1390
1391
1392
1393
                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
1394
        else:
lintangsutawika's avatar
lintangsutawika committed
1395
1396
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1397
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1398
            )
1399
1400
1401

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1402
    def aggregation(self) -> dict:
1403
        return {k.name: k.aggregation_fn for k in self.config._metric_list}
1404

Baber Abbasi's avatar
Baber Abbasi committed
1405
    def higher_is_better(self) -> dict:
1406
        return {k.name: k.higher_is_better for k in self.config._metric_list}
1407

Baber Abbasi's avatar
Baber Abbasi committed
1408
1409
1410
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1411
    @property
Baber's avatar
cleanup  
Baber committed
1412
    def task_name(self) -> str | None:
Lintang Sutawika's avatar
Lintang Sutawika committed
1413
1414
        return getattr(self.config, "task", None)

1415
1416
1417
1418
1419
    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
1420
            f"num_samples={len(self.eval_docs)})"
1421
1422
        )

1423
1424

class MultipleChoiceTask(Task):
1425
    OUTPUT_TYPE = "loglikelihood"
1426

baberabb's avatar
baberabb committed
1427
    def doc_to_target(self, doc: dict) -> str:
1428
1429
        return " " + doc["choices"][doc["gold"]]

Baber's avatar
cleanup  
Baber committed
1430
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> list[Instance]:
1431
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1432
1433
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1434
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1435
                doc=doc,
1436
                arguments=(ctx, " {}".format(choice)),
1437
                idx=i,
1438
1439
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1440
1441
            for i, choice in enumerate(doc["choices"])
        ]
1442

Baber's avatar
cleanup  
Baber committed
1443
    def process_results(self, doc: dict, results: Iterable[tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1444
1445
1446
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
        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
1458
    def higher_is_better(self) -> dict:
1459
1460
1461
1462
1463
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1464
    def aggregation(self) -> dict:
1465
1466
1467
1468
1469
1470
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1471
class PerplexityTask(Task):
1472
1473
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1474
    def has_training_docs(self) -> bool:
1475
1476
        return False

Baber's avatar
cleanup  
Baber committed
1477
    def fewshot_examples(self, k: int, rnd) -> list:
1478
1479
1480
1481
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1482
1483
        return []

baberabb's avatar
baberabb committed
1484
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1485
1486
1487
1488
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1489
1490
1491

        return ""

baberabb's avatar
baberabb committed
1492
    def higher_is_better(self) -> dict:
1493
1494
1495
1496
1497
1498
1499
1500
1501
        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
1502
    def doc_to_text(self, doc) -> str:
1503
1504
1505
1506
1507
        return ""

    def doc_to_target(self, doc):
        return doc

Baber's avatar
cleanup  
Baber committed
1508
    def construct_requests(self, doc: dict, ctx: str | None, **kwargs):
1509
1510
        if bool(ctx):
            raise ValueError
1511

lintangsutawika's avatar
lintangsutawika committed
1512
1513
1514
1515
1516
1517
1518
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1519

Baber's avatar
cleanup  
Baber committed
1520
    def process_results(self, doc: dict, results: tuple[float]) -> dict:
1521
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1522
1523
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1524
1525
1526
1527
1528
1529
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1530
    def aggregation(self) -> dict:
1531
1532
1533
1534
1535
1536
1537
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1538
    def count_bytes(cls, doc) -> int:
1539
1540
1541
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1542
    def count_words(cls, doc) -> int:
1543
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1544
        return len(re.split(r"\s+", doc))