"vscode:/vscode.git/clone" did not exist on "70cf4abccc03d69d35894ab8bec3e75bef48f00f"
task.py 59.7 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
11
from functools import cached_property
from types import MethodType
Baber's avatar
Baber committed
12
from typing import TYPE_CHECKING, Any, Literal, overload
13
14
15

import datasets
import numpy as np
16
from tqdm import tqdm
Baber's avatar
Baber committed
17
from typing_extensions import deprecated
18
19

from lm_eval import utils
20
21
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
22
from lm_eval.api.utils import check_gold_index_error
23
from lm_eval.caching.cache import load_from_cache, save_to_cache
Baber's avatar
Baber committed
24
from lm_eval.config.metric import MetricConfig
Baber's avatar
Baber committed
25
from lm_eval.config.task import DataSet, TaskConfig
26
27
from lm_eval.filters import build_filter_ensemble

28

29
30
31
32
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
33
    "generate_until",
34
35
]

Baber's avatar
cleanup  
Baber committed
36
if TYPE_CHECKING:
Baber's avatar
Baber committed
37
    pass
Baber's avatar
cleanup  
Baber committed
38
39


Lintang Sutawika's avatar
Lintang Sutawika committed
40
eval_logger = logging.getLogger(__name__)
41

lintangsutawika's avatar
lintangsutawika committed
42

43
44
45
46
47
48
49
50
51
52
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
53
    VERSION: int | str | None = None
54

55
56
    # 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
57
    DATASET_PATH: str | None = None
58
59

    # The name of a subset within `DATASET_PATH`.
Baber's avatar
cleanup  
Baber committed
60
    DATASET_NAME: str | None = None
61

Baber's avatar
cleanup  
Baber committed
62
    OUTPUT_TYPE: OutputType | None = None
lintangsutawika's avatar
lintangsutawika committed
63

64
65
    def __init__(
        self,
Baber's avatar
cleanup  
Baber committed
66
67
68
69
        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
70
    ) -> None:
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
        """
        :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
93
94
95
        self._training_docs: list | None = None
        self._fewshot_docs: list | None = None
        self._instances: list[Instance] | None = None
96

97
        self._config: TaskConfig = TaskConfig.from_yaml({**config})
98

99
        self._filters = [build_filter_ensemble("none", [("take_first", None)])]
Baber's avatar
cleanup  
Baber committed
100
        self.fewshot_rnd: random.Random | None = (
101
102
            None  # purposely induce errors in case of improper usage
        )
103

104
105
    def download(
        self,
Baber's avatar
cleanup  
Baber committed
106
107
        data_dir: str | None = None,
        cache_dir: str | None = None,
108
109
        download_mode=None,
    ) -> None:
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
        """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.
        """
Baber's avatar
Baber committed
134
        assert self.DATASET_PATH is not None, "DATASET_PATH must be set in Task class"
135
136
137
138
139
140
141
        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,
        )
142

143
    @property
144
    def config(self) -> TaskConfig:
145
146
147
        """Returns the TaskConfig associated with this class."""
        return self._config

148
    @property
Baber's avatar
Baber committed
149
    def has_training_docs(self) -> bool:
150
        """Whether the task has a training set"""
Baber's avatar
Baber committed
151
        raise NotImplementedError
152

153
    @property
Baber's avatar
Baber committed
154
    def has_validation_docs(self) -> bool:
155
        """Whether the task has a validation set"""
Baber's avatar
Baber committed
156
        raise NotImplementedError
157

158
    @property
Baber's avatar
Baber committed
159
    def has_test_docs(self) -> bool:
160
        """Whether the task has a test set"""
Baber's avatar
Baber committed
161
        raise NotImplementedError
162

Baber's avatar
Baber committed
163
    def training_docs(self) -> DataSet | None:
164
165
166
167
168
169
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

Baber's avatar
Baber committed
170
    def validation_docs(self) -> DataSet | None:
171
172
173
174
175
176
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

Baber's avatar
Baber committed
177
    def test_docs(self) -> DataSet | None:
178
179
180
181
182
183
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

Baber's avatar
Baber committed
184
    def fewshot_docs(self) -> DataSet | None:
185
186
187
188
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
189
        if self.has_training_docs:
190
            return self.training_docs()
191
        elif self.has_validation_docs:
192
193
            return self.validation_docs()
        else:
Baber's avatar
Baber committed
194
            if self.config.num_fewshot and self.config.num_fewshot > 0:
Baber Abbasi's avatar
Baber Abbasi committed
195
196
197
198
                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."
                )
199
200
            return self.test_docs()

201
    def _process_doc(self, doc: dict) -> dict:
202
203
204
205
206
207
208
209
210
        """
        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
211

212
    @property
Baber's avatar
cleanup  
Baber committed
213
    def instances(self) -> list[Instance]:
214
215
216
217
218
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

219
    def fewshot_examples(self, k: int, rnd) -> Iterable[dict]:
220
221
222
223
224
        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
225
    def doc_to_decontamination_query(self, doc: dict):
226
        raise NotImplementedError(
227
228
229
230
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
231
    def doc_to_text(self, doc: dict) -> str:
232
233
234
        pass

    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
235
    def doc_to_target(self, doc: dict) -> str | int:
236
237
        pass

238
    # not an abstractmethod because not every language-only task has to implement this
Baber's avatar
cleanup  
Baber committed
239
    def doc_to_image(self, doc: dict):
240
241
        raise NotImplementedError

Baber's avatar
cleanup  
Baber committed
242
    def doc_to_audio(self, doc: dict):
243
244
        raise NotImplementedError

Baber's avatar
cleanup  
Baber committed
245
    def doc_to_prefix(self, doc: dict) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
246
247
        return ""

248
249
    def build_all_requests(
        self,
250
        *,
Baber's avatar
cleanup  
Baber committed
251
252
        limit: int | None = None,
        samples: list[int] | None = None,
253
254
255
256
        rank: int = 0,
        world_size: int = 1,
        cache_requests: bool = False,
        rewrite_requests_cache: bool = False,
Baber's avatar
cleanup  
Baber committed
257
        system_instruction: str | None = None,
258
259
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
260
        chat_template: Callable | None = None,
261
        tokenizer_name: str = "",
262
    ) -> None:
263
        """Build a set of Instances for a task, and store them in task.instances"""
264
265
266
267

        # used with caching
        og_limit = limit

268
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
269
270
271
272
273
274
275
        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 ""
        )
276
        cache_key += f"-tokenizer{tokenizer_name}"
277

Baber Abbasi's avatar
Baber Abbasi committed
278
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
279
280
281
282
283
284
285
286
287
288
289
290
291

        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
292
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
293

294
        instances = []
295
296
297
298
299
300
301
302
303
304

        # 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(
305
306
307
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
308
309
310
311
312
313
314
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
315
        ):
316
            # sample fewshot context #TODO: need to offset doc_id by rank now!
317
            fewshot_ctx = self.fewshot_context(
318
                doc,
319
320
321
322
323
324
325
                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
326
                gen_prefix=self.doc_to_prefix(doc),
327
            )
328

329
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
330
331
332
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
Baber's avatar
Baber committed
333
                metadata=(self.config.task, doc_id, self.config.repeats),
334
                apply_chat_template=apply_chat_template,
335
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
336
            )
337
338
339
340

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

341
342
343
344
345
346
347
348
349
350
351
352
353
            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
354

355
356
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
357

358
359
360
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

361
    @abc.abstractmethod
Baber's avatar
cleanup  
Baber committed
362
    def construct_requests(self, doc: dict, ctx: list[dict] | str, **kwargs):
363
364
365
366
367
368
369
370
371
372
373
374
375
376
        """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
377
            The number of times each instance in a dataset is inferred on. Defaults to 1,
378
379
380
381
382
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
383
    def process_results(self, doc: dict, results: list) -> dict[str, Any]:
384
385
386
387
388
389
390
391
392
        """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.
        """
393
        raise NotImplementedError
394

Baber's avatar
Baber committed
395
    @deprecated("not used anymore")
396
397
398
399
400
401
    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
402
        return True
403

Baber's avatar
Baber committed
404
    @deprecated("not used anymore")
405
406
407
408
409
410
    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
411
        return True
412

413
414
415
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
416
    @classmethod
Baber's avatar
Baber committed
417
    def count_bytes(cls, doc: str) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
418
419
420
421
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
422
    def count_words(cls, doc: str) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
423
424
425
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

426
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
427
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
428
429
430
431
432
433
434
        """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
435
436
437
438
439
        :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.
440
441
442
        :returns: str
            The fewshot context.
        """
443
        if rnd is None:
444
445
446
447
448
449
            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
450

451
        description = description if description else ""
452
453

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
454
            labeled_examples = ""
455
        else:
lintangsutawika's avatar
lintangsutawika committed
456
            # for sets with no training docs, draw from other set *but ensure no overlap with current doc*
457
            if self.has_training_docs:
lintangsutawika's avatar
lintangsutawika committed
458
459
460
461
462
                fewshotex = self.fewshot_examples(k=num_fewshot, rnd=rnd)
            else:
                if self._fewshot_docs is None:
                    self._fewshot_docs = list(
                        self.validation_docs()
463
                        if self.has_validation_docs
lintangsutawika's avatar
lintangsutawika committed
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
                        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
480
            )
481
482

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
483
        return description + labeled_examples + example
484

Baber's avatar
cleanup  
Baber committed
485
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
486
        """Iterates over FilterEnsembles and applies them to instances"""
Baber's avatar
cleanup  
Baber committed
487
        if hasattr(self, "_filters") and self._instances:
lintangsutawika's avatar
lintangsutawika committed
488
            for f in self._filters:
489
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
490
        else:
Baber's avatar
cleanup  
Baber committed
491
492
493
            eval_logger.warning(
                "No filter defined or no instances, passing through instances"
            )
lintangsutawika's avatar
lintangsutawika committed
494
            return self._instances
495

baberabb's avatar
baberabb committed
496
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
497
        """Returns the config as a dictionary."""
498
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
499
        # (num_fewshot)
500
        return self.config.to_dict()
501

Baber Abbasi's avatar
Baber Abbasi committed
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
    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
521
522
523
524
525
526
527
        # 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
528

Baber's avatar
cleanup  
Baber committed
529
    def set_fewshot_seed(self, seed: int | None = None) -> None:
530
531
532
533
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

534
    @property
Baber's avatar
cleanup  
Baber committed
535
    def eval_docs(self) -> datasets.Dataset | Iterable[dict]:
536
        if self.has_test_docs:
537
            return self.test_docs()
538
        elif self.has_validation_docs:
539
540
            return self.validation_docs()
        else:
541
542
543
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
544
545

    def doc_iterator(
546
547
548
        self,
        *,
        rank: int = 0,
Baber's avatar
cleanup  
Baber committed
549
        limit: int | None = None,
550
        world_size: int = 1,
Baber's avatar
cleanup  
Baber committed
551
552
        samples: list[int] | None = None,
    ) -> Iterator[tuple[int, Any]]:
553
554
        if samples:
            n = len(self.eval_docs)
Baber's avatar
cleanup  
Baber committed
555
            assert all(e < n for e in samples), (
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
                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),
            )
575
576
        return doc_iterator

577
578

class ConfigurableTask(Task):
579
    VERSION = "Yaml"
580
    OUTPUT_TYPE = None
581
    CONFIG = None
582
583

    def __init__(
584
585
586
587
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
Baber's avatar
Baber committed
588
        config: Mapping[str, Any] | None = None,
Baber's avatar
Baber committed
589
    ) -> None:
590
        # Get pre-configured attributes
591
        self._config = self.CONFIG
592

593
        # Use new configurations if there was no preconfiguration
594
        if self.config is None:
595
            self._config = TaskConfig.from_yaml(config)
596
597
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
598
            if config is not None:
599
                self._config.__dict__.update(config)
600

601
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
602
603
604
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
605

Baber's avatar
cleanup  
Baber committed
606
607
        if isinstance(self.config.metadata, dict) and "version" in self.config.metadata:
            self.VERSION = self.config.metadata["version"]
608

609
        if self.config.output_type is not None:
610
611
612
613
            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)}'"
                )
614
            self.OUTPUT_TYPE = self.config.output_type
615

616
617
618
619
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

620
621
622
623
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
624
625
626
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

627
628
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
629

630
631
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
632

633
        # self.metric_list: list[MetricConfig] = self.config.get_metrics
634

635
        self.download(self.config.dataset_kwargs)
636
637
638
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
639
        self._filters = self.config.get_filters
Baber's avatar
Baber committed
640

Baber's avatar
Baber committed
641
642
643
644
645
646
647
        # 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
648

649
650
651
652
        if (
            self.config.fewshot_cfg.num_fewshot() > 0
            and self.fewshot_docs() is not None
        ):
Baber's avatar
Baber committed
653
654
655
            self.fewshot_rnd = random.Random()
            self.sampler = self.config.fewshot_cfg.init_sampler(
                list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
656
            )
657
        self.task_docs = self.eval_docs
658

659

Baber Abbasi's avatar
Baber Abbasi committed
660
    def download(
Baber's avatar
cleanup  
Baber committed
661
        self, dataset_kwargs:dict[str, Any] | None = None, **kwargs
Baber Abbasi's avatar
Baber Abbasi committed
662
    ) -> None:
Baber Abbasi's avatar
Baber Abbasi committed
663
664
        from packaging.version import parse as vparse

665
666
667
668
        self.config.dataset_kwargs, self.config.metadata = (
            self.config.dataset_kwargs or {},
            self.config.metadata or {},
        )
Baber Abbasi's avatar
Baber Abbasi committed
669
670
        if dataset_kwargs and vparse(datasets.__version__) >= vparse("4.0.0"):
            dataset_kwargs.pop("trust_remote_code", None)
671
        if isinstance(df := self.config.custom_dataset, Callable):
Baber Abbasi's avatar
Baber Abbasi committed
672
673
674
675
            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."
            )
676
            self.dataset = df(**(self.config.dataset_kwargs | self.config.metadata))
Baber Abbasi's avatar
Baber Abbasi committed
677
        else:
Baber's avatar
Baber committed
678
679
680
            assert self.config.dataset_path is not None, (
                "dataset_path must be set in TaskConfig"
            )
Baber Abbasi's avatar
Baber Abbasi committed
681
            self.dataset = datasets.load_dataset(
682
683
684
                path=self.config.dataset_path,
                name=self.config.dataset_name,
                **self.config.dataset_kwargs,
Baber Abbasi's avatar
Baber Abbasi committed
685
            )
686

687
    @cached_property
baberabb's avatar
baberabb committed
688
    def has_training_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
689
        return self.config.training_split is not None
690

691
    @cached_property
baberabb's avatar
baberabb committed
692
    def has_validation_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
693
        return self.config.validation_split is not None
694

695
    @cached_property
baberabb's avatar
baberabb committed
696
    def has_test_docs(self) -> bool:
Baber's avatar
cleanup  
Baber committed
697
        return self.config.test_split is not None
698

Baber's avatar
Baber committed
699
    def training_docs(self) -> DataSet | None:
700
        if self.has_training_docs:
701
702
703
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
704
                )
705
            return self.dataset[self.config.training_split]
706

Baber's avatar
Baber committed
707
    def validation_docs(self) -> DataSet | None:
708
        if self.has_validation_docs:
709
710
711
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
712
                )
713
            return self.dataset[self.config.validation_split]
714

Baber's avatar
Baber committed
715
    def test_docs(self) -> DataSet | None:
716
        if self.has_test_docs:
717
718
719
            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]
720

721
    def fewshot_docs(self):
Baber's avatar
Baber committed
722
723
724
725
726
727
        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
728
729
730
731
732
733
734
735
736
737
        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
738
739

        return super().fewshot_docs()
740

KonradSzafer's avatar
KonradSzafer committed
741
742
    @staticmethod
    def append_target_question(
Baber's avatar
cleanup  
Baber committed
743
        labeled_examples: list[dict[str, str]],
KonradSzafer's avatar
KonradSzafer committed
744
745
        question: str,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
746
        gen_prefix: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
747
748
749
750
751
752
753
754
    ) -> 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":
755
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
756
757
            # if last message is user, append to it to avoid two user messages in a row
            else:
758
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
759
760
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
761
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
762
763
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
764

lintangsutawika's avatar
lintangsutawika committed
765
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
766
767
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
768
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
769
        num_fewshot: int,
Baber's avatar
cleanup  
Baber committed
770
        system_instruction: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
771
772
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
cleanup  
Baber committed
773
774
775
        chat_template: Callable | None = None,
        gen_prefix: str | None = None,
    ) -> str | list[str] | None:
lintangsutawika's avatar
lintangsutawika committed
776
777
778
779
780
781
782
        """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
783
784
785
786
787
788
        :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.
789
790
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
791
792
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
793
794
795
        :returns: str
            The fewshot context.
        """
Baber's avatar
cleanup  
Baber committed
796
        labeled_examples = [] if apply_chat_template else ""
KonradSzafer's avatar
KonradSzafer committed
797
798

        # get task description
799
800
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
801

KonradSzafer's avatar
KonradSzafer committed
802
803
804
805
806
807
808
809
810
        # 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
811
        else:
KonradSzafer's avatar
KonradSzafer committed
812
813
814
815
816
817
818
819
820
821
822
823
824
            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
825
826
827
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
828
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
829
830
831
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
832
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
833
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
834
                )
lintangsutawika's avatar
lintangsutawika committed
835
836

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
837
838
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
839
                # TODO: append prefill?
840
841
                if not labeled_examples:
                    return ""
842
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
843
844
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
845
846
847
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
848
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
849
850
851
852
853
854
855
                )
            # 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
856
857
858
859
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
860
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
861
862
863
864
865
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber's avatar
cleanup  
Baber committed
866
                            add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
867
868
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
869
870
871
872
873
874
                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
875
876
877
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
878
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
879
880
881
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
882
883
884
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
885
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
886
887
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
888
889
            return chat_template(
                labeled_examples,
Baber's avatar
cleanup  
Baber committed
890
                add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
891
            )
892
        else:
Baber Abbasi's avatar
Baber Abbasi committed
893
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
894
895
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
896
897
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
898
899
            if self.multiple_input:
                return labeled_examples
900
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
901
                return labeled_examples + example + prefix
902
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
903
                return [labeled_examples + ex + prefix for ex in example]
904
905
906
            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
907
                    return labeled_examples + choices[example] + prefix
908
                else:
Baber Abbasi's avatar
Baber Abbasi committed
909
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
910

Baber's avatar
cleanup  
Baber committed
911
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
912
        """Iterates over FilterEnsembles and applies them to instances"""
913
        if hasattr(self, "_filters") and self._instances:
914
            for f in self._filters:
915
                f.ensemble.apply(self._instances)
916
        else:
917
918
919
            eval_logger.warning(
                "No filter defined or instances found. Passing through instances"
            )
920
921
            return self._instances

922
    def should_decontaminate(self):
923
        return self.config.should_decontaminate
924

Baber Abbasi's avatar
Baber Abbasi committed
925
    def doc_to_decontamination_query(self, doc: dict):
926
        if self.config.should_decontaminate:
927
928
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
929
            else:
930
931
932
933
934
935
936
937
938
939
940
                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
                        )
                    )
941

942
    def _process_doc(self, doc: dict) -> dict:
943
944
945
946
947
948
949
950
951
952
        """
        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

Baber's avatar
Baber committed
953
954
955
956
957
958
959
960
961
962
963
964
    @overload
    def doc_to_text(self, doc: dict, doc_to_text: None = None) -> str | int: ...

    @overload
    def doc_to_text(self, doc: dict, doc_to_text: int) -> int: ...

    @overload
    def doc_to_text(self, doc: dict, doc_to_text: str) -> str: ...

    @overload
    def doc_to_text(self, doc: dict, doc_to_text: Callable[..., str]) -> str: ...

965
966
    def doc_to_text(
        self, doc: dict, doc_to_text: int | str | Callable[..., str] | None = None
Baber's avatar
Baber committed
967
    ) -> str | int:
Baber's avatar
Baber committed
968
969
        # if self.prompt is not None:
        #     doc_to_text = self.prompt
970
        doc_to_text = doc_to_text or self.config.doc_to_text
971
972
        if doc_to_text in doc:
            return doc[doc_to_text]
973
        elif isinstance(doc_to_text, str):
974
975
976
            text_string = utils.apply_template(doc_to_text, doc)
            if text_string.isdigit() and self.config.doc_to_choice is not None:
                return ast.literal_eval(text_string)
977
            else:
978
979
980
                return text_string
        elif isinstance(doc_to_text, int):
            return doc_to_text
lintangsutawika's avatar
lintangsutawika committed
981
        # Used when applying a Promptsource template
Baber's avatar
Baber committed
982
983
984
985
986
987
988
        # 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
989
        else:
990
            print(type(doc_to_text))
991
            raise TypeError
992

Baber's avatar
Baber committed
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
    @overload
    def doc_to_target(
        self, doc: dict, doc_to_target: None = None
    ) -> int | str | list[int]: ...

    @overload
    def doc_to_target(self, doc: dict, doc_to_target: int) -> int: ...

    @overload
    def doc_to_target(self, doc: dict, doc_to_target: str) -> int | str | list[int]: ...

    @overload
    def doc_to_target(self, doc: dict, doc_to_target: list) -> list[int]: ...

    @overload
    def doc_to_target(
        self, doc: dict, doc_to_target: Callable[..., int | str | list[int]]
    ) -> int | str | list[int]: ...

Baber's avatar
cleanup  
Baber committed
1012
    def doc_to_target(self, doc: dict, doc_to_target=None) -> int | str | list[int]:
Baber's avatar
Baber committed
1013
1014
        # if self.prompt is not None:
        #     doc_to_target = self.prompt
1015
1016
1017
        doc_to_target = doc_to_target or self.config.doc_to_target
        if doc_to_target in doc:
            return doc[doc_to_target]
1018
        elif isinstance(doc_to_target, str):
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
            target_string = utils.apply_template(doc_to_target, doc)
            if target_string.isdigit() and self.config.doc_to_choice is not None:
                return ast.literal_eval(target_string)
            # elif (
            #     len(target_string) >= 2
            #     and (target_string[0] == "[")
            #     and (target_string[-1] == "]")
            # ):
            #     try:
            #         return ast.literal_eval(target_string)
            #     except (SyntaxError, ValueError):
            #         return target_string
1031
            else:
1032
1033
1034
                return target_string

        elif isinstance(doc_to_target, (int, list)):
1035
            return doc_to_target
1036
1037
1038
1039
        # elif isinstance(doc_to_target, list):
        #     return doc_to_target
        # elif callable(doc_to_target):
        #     return doc_to_target(doc)
Baber's avatar
Baber committed
1040
1041
1042
1043
1044
1045
1046
1047
        # # 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
1048
1049
        else:
            raise TypeError
1050

Baber's avatar
Baber committed
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: None = None) -> list[str]: ...

    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: str) -> list[str]: ...

    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: list) -> list[str]: ...

    @overload
    def doc_to_choice(self, doc: dict, doc_to_choice: dict) -> list[str]: ...

    @overload
    def doc_to_choice(
        self, doc: dict, doc_to_choice: Callable[..., list[str]]
    ) -> list[str]: ...

Baber's avatar
cleanup  
Baber committed
1068
    def doc_to_choice(
Baber's avatar
Baber committed
1069
1070
        self,
        doc: dict,
Baber's avatar
cleanup  
Baber committed
1071
1072
        doc_to_choice: str | list | dict | Callable[..., list[str]] | None = None,
    ) -> list[str]:
Baber's avatar
Baber committed
1073
1074
1075
        # 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
1076
            doc_to_choice = doc_to_choice
1077
        elif self.config.doc_to_choice is None:
1078
            eval_logger.error("doc_to_choice was called but not set in config")
Baber's avatar
Baber committed
1079
            doc_to_choice = None
1080
        else:
1081
            doc_to_choice = self.config.doc_to_choice
1082

1083
        if isinstance(doc_to_choice, str):
1084
            if doc_to_choice in doc:
1085
1086
1087
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1088
        elif isinstance(doc_to_choice, list):
1089
            return doc_to_choice
1090
1091
        # elif isinstance(doc_to_choice, dict):
        #     return list(doc_to_choice.values())
Baber's avatar
Baber committed
1092
1093
        # elif hasattr(doc_to_choice, "get_answer_choices_list"):
        #     return doc_to_choice.get_answer_choices_list(doc)
1094
1095
        else:
            raise TypeError
1096

Baber's avatar
Baber committed
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
    @overload
    def doc_to_image(self, doc: dict, doc_to_image: None = None) -> None: ...

    @overload
    def doc_to_image(self, doc: dict, doc_to_image: list) -> list: ...

    @overload
    def doc_to_image(self, doc: dict, doc_to_image: str) -> int | str | None: ...

    @overload
    def doc_to_image(self, doc: dict, doc_to_image: Callable[..., Any]) -> Any: ...

Baber's avatar
cleanup  
Baber committed
1109
    def doc_to_image(self, doc: dict, doc_to_image=None) -> int | str | list | None:
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
        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
Baber committed
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: None = None) -> None: ...

    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: list) -> list: ...

    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: str) -> int | str | None: ...

    @overload
    def doc_to_audio(self, doc: Any, doc_to_audio: Callable[..., Any]) -> Any: ...

Baber's avatar
cleanup  
Baber committed
1144
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> int | str | list | None:
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
        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
1167
    def doc_to_prefix(self, doc: dict) -> str | None:
Baber Abbasi's avatar
Baber Abbasi committed
1168
        if (gen_prefix := self.config.gen_prefix) is not None:
1169
            if gen_prefix in doc:
Baber Abbasi's avatar
Baber Abbasi committed
1170
1171
1172
1173
1174
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1175
1176
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
Baber's avatar
cleanup  
Baber committed
1177
    ) -> list[Instance] | Instance:
1178
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1179
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1180

1181
1182
        aux_arguments = None

1183
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1184
            arguments = (ctx, self.doc_to_target(doc))
1185
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1186
            arguments = (self.doc_to_target(doc),)
1187
        elif self.OUTPUT_TYPE == "multiple_choice":
1188
            choices = self.doc_to_choice(doc)
1189
            target_delimiter = self.config.target_delimiter
1190
1191
            if apply_chat_template:
                target_delimiter = ""
1192
1193
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1194
                # apply chat_template to choices if apply_chat_template
1195
                cont = self.doc_to_target(doc)
1196

1197
                arguments = [
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1208
                ]
1209
            else:
1210
                # Otherwise they are placed in the continuation
1211
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1212

1213
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
1214
            if "acc_mutual_info" in [m.metric_name for m in self.config._metric_list]:
1215
1216
1217
1218
1219
1220
                # 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.
1221
1222
1223
1224
                # 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
                ]
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239

                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)},
            }

1240
1241
1242
1243
1244
1245
1246
1247
        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)},
            }

1248
1249
1250
1251
1252
1253
1254
        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":
1255
            request_list = [
1256
1257
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1258
                    doc=doc,
1259
                    arguments=arg,
1260
                    idx=i,
1261
1262
                    **kwargs,
                )
1263
                for i, arg in enumerate(arguments)
1264
            ]
1265
1266

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1267

lintangsutawika's avatar
lintangsutawika committed
1268
        return Instance(
1269
1270
1271
1272
1273
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1274
        )
1275

1276
    def process_results(self, doc: dict, results: list) -> dict[str, Any]:
1277
        result_dict = {}
1278
        use_metric = list(m.metric_name for m in self.config._metric_list)
1279
1280
1281
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1282
1283
1284
1285
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1286
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
1287
1288
1289
1290
1291
1292
            (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
1293
            return {
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
                **(
                    {"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
1309
            }
1310
        elif self.OUTPUT_TYPE == "multiple_choice":
1311
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1312

Baber's avatar
cleanup  
Baber committed
1313
            # retrieve choices in list[str] form, to compute choice lengths, etc.
1314
            choices = self.doc_to_choice(doc)
1315
1316
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1317
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1318
1319
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1320
1321
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1322
1323
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1324
                # and this stores our "regular" conditional loglikelihoods
1325
                lls = lls[: len(choices)]
Baber's avatar
Baber committed
1326
1327
            else:
                lls_unconditional = None
1328

1329
1330
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1331

1332
1333
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1334
            else:
1335
                gold = self.doc_to_target(doc)
1336

1337
            gold, gold_index_error = check_gold_index_error(choices, gold)
1338
1339
1340

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1341
                    f"Label index was not in within range of available choices,"
1342
1343
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1344

1345
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1346
1347
                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
1348
                exact_match = int(any(is_greedy[i] if i != -100 else 0 for i in gold))
lintangsutawika's avatar
lintangsutawika committed
1349
1350
1351
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1352
                # 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
1353
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1354

Lintang Sutawika's avatar
Lintang Sutawika committed
1355
1356
1357
1358
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1359
            result_dict = {
1360
                **({"acc": acc} if "acc" in use_metric else {}),
1361
1362
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1363
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1364
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1365
1366
1367
1368
1369
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1370
1371
            }

1372
            if "acc_mutual_info" in use_metric:
Baber's avatar
Baber committed
1373
1374
1375
                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
1376
1377
1378
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1379
1380
1381
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1382
        elif self.OUTPUT_TYPE == "generate_until":
1383
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1384
            result = results[0]
Baber's avatar
Baber committed
1385
            for metric in self.config._metric_list:
1386
                try:
Baber's avatar
Baber committed
1387
                    result_score = metric.fn(
1388
1389
                        references=[gold] if not isinstance(gold, list) else gold,
                        predictions=[result],
Baber's avatar
Baber committed
1390
                        **metric.kwargs,
1391
1392
                    )
                except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
Baber's avatar
Baber committed
1393
                    result_score = metric.fn([gold, result])
1394
1395
1396
1397
1398
1399
1400
                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
1401
        else:
lintangsutawika's avatar
lintangsutawika committed
1402
1403
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1404
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1405
            )
1406
1407
1408

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1409
    def aggregation(self) -> dict:
1410
        return {k.name: k.aggregation_fn for k in self.config._metric_list}
1411

Baber Abbasi's avatar
Baber Abbasi committed
1412
    def higher_is_better(self) -> dict:
1413
        return {k.name: k.higher_is_better for k in self.config._metric_list}
1414

Baber Abbasi's avatar
Baber Abbasi committed
1415
1416
1417
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1418
    @property
Baber's avatar
cleanup  
Baber committed
1419
    def task_name(self) -> str | None:
Lintang Sutawika's avatar
Lintang Sutawika committed
1420
1421
        return getattr(self.config, "task", None)

Baber's avatar
Baber committed
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
    def runtime_checks(self, test_doc):
        # Test One Doc
        self.features: list[str] = list(self.task_docs.features.keys())
        self.multiple_target = 0
        test_text = self.doc_to_text(test_doc)
        test_target = self.doc_to_target(test_doc)

        if self.config.doc_to_choice is not None:
            test_choice = self.doc_to_choice(test_doc)
            if not isinstance(test_choice, list):
                eval_logger.error("doc_to_choice must return list")
            # else:
            #     num_choice = len(test_choice)

            if isinstance(test_text, int):
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
        else:
            test_choice = None

        if isinstance(test_target, list):
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
            self.multiple_target = len(test_target)
        else:
            if (isinstance(test_target, int)) and (test_choice is not None):
                test_target = test_choice[test_target]
            else:
                test_target = str(test_target)

        check_choices = test_choice if test_choice is not None else [test_target]
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = choice[0].isspace()
                delimiter_has_whitespace = (
                    self.config.target_delimiter.rstrip()
                    != self.config.target_delimiter
                )

                if delimiter_has_whitespace and choice_has_whitespace:
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
                    eval_logger.debug(
                        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'
                    )

1472
1473
1474
1475
1476
    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
1477
            f"num_samples={len(self.eval_docs)})"
1478
1479
        )

1480
1481

class MultipleChoiceTask(Task):
1482
    OUTPUT_TYPE = "loglikelihood"
1483

baberabb's avatar
baberabb committed
1484
    def doc_to_target(self, doc: dict) -> str:
1485
1486
        return " " + doc["choices"][doc["gold"]]

Baber's avatar
cleanup  
Baber committed
1487
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> list[Instance]:
1488
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1489
1490
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1491
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1492
                doc=doc,
Baber's avatar
Baber committed
1493
                arguments=(ctx, f" {choice}"),
1494
                idx=i,
1495
1496
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1497
1498
            for i, choice in enumerate(doc["choices"])
        ]
1499

Baber's avatar
cleanup  
Baber committed
1500
    def process_results(self, doc: dict, results: Iterable[tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1501
1502
1503
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
        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
1515
    def higher_is_better(self) -> dict:
1516
1517
1518
1519
1520
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1521
    def aggregation(self) -> dict:
1522
1523
1524
1525
1526
1527
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1528
class PerplexityTask(Task):
1529
1530
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1531
    def has_training_docs(self) -> bool:
1532
1533
        return False

Baber's avatar
cleanup  
Baber committed
1534
    def fewshot_examples(self, k: int, rnd) -> list:
1535
1536
1537
1538
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1539
1540
        return []

baberabb's avatar
baberabb committed
1541
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1542
1543
1544
1545
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1546
1547
1548

        return ""

baberabb's avatar
baberabb committed
1549
    def higher_is_better(self) -> dict:
1550
1551
1552
1553
1554
1555
1556
1557
1558
        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
1559
    def doc_to_text(self, doc) -> str:
1560
1561
1562
1563
1564
        return ""

    def doc_to_target(self, doc):
        return doc

Baber's avatar
cleanup  
Baber committed
1565
    def construct_requests(self, doc: dict, ctx: str | None, **kwargs):
1566
1567
        if bool(ctx):
            raise ValueError
1568

lintangsutawika's avatar
lintangsutawika committed
1569
1570
1571
1572
1573
1574
1575
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1576

Baber's avatar
cleanup  
Baber committed
1577
    def process_results(self, doc: dict, results: tuple[float]) -> dict:
1578
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1579
1580
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1581
1582
1583
1584
1585
1586
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1587
    def aggregation(self) -> dict:
1588
1589
1590
1591
1592
1593
1594
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1595
    def count_bytes(cls, doc) -> int:
1596
1597
1598
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1599
    def count_words(cls, doc) -> int:
1600
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1601
        return len(re.split(r"\s+", doc))