task.py 58.4 KB
Newer Older
Baber's avatar
nit  
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
import re
8
from collections.abc import Callable, Iterable, Iterator, Mapping
9
from copy import deepcopy
10
from typing import (
Baber's avatar
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.caching.cache import load_from_cache, save_to_cache
Baber's avatar
Baber committed
25
26
from lm_eval.config.metric import MetricConfig
from lm_eval.config.task import TaskConfig
27
28
from lm_eval.filters import build_filter_ensemble

29

30
31
32
33
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
34
    "generate_until",
35
36
]

Baber's avatar
Baber committed
37
if TYPE_CHECKING:
Baber's avatar
Baber committed
38
    pass
Baber's avatar
Baber committed
39
40


Lintang Sutawika's avatar
Lintang Sutawika committed
41
eval_logger = logging.getLogger(__name__)
42

lintangsutawika's avatar
lintangsutawika committed
43

44
45
46
47
48
49
50
51
52
53
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
nit  
Baber committed
54
    VERSION: int | str | None = None
55

56
57
    # 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
nit  
Baber committed
58
    DATASET_PATH: str | None = None
59
60

    # The name of a subset within `DATASET_PATH`.
Baber's avatar
nit  
Baber committed
61
    DATASET_NAME: str | None = None
62

Baber's avatar
nit  
Baber committed
63
    OUTPUT_TYPE: OutputType | None = None
lintangsutawika's avatar
lintangsutawika committed
64

65
66
    def __init__(
        self,
Baber's avatar
nit  
Baber committed
67
68
69
70
        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
71
    ) -> None:
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
        """
        :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
nit  
Baber committed
94
95
96
        self._training_docs: list | None = None
        self._fewshot_docs: list | None = None
        self._instances: list[Instance] | None = None
97

98
        self._config: TaskConfig = TaskConfig.from_yaml({**config})
99

100
        self._filters = [build_filter_ensemble("none", [("take_first", None)])]
Baber's avatar
nit  
Baber committed
101
        self.fewshot_rnd: random.Random | None = (
102
103
            None  # purposely induce errors in case of improper usage
        )
104

105
106
    def download(
        self,
Baber's avatar
nit  
Baber committed
107
108
        data_dir: str | None = None,
        cache_dir: str | None = None,
109
110
        download_mode=None,
    ) -> None:
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
        """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.
        """
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
    @abc.abstractmethod
Baber's avatar
Baber committed
149
    def has_training_docs(self) -> bool:
150
151
152
153
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
154
    def has_validation_docs(self) -> bool:
155
156
157
158
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
159
    def has_test_docs(self) -> bool:
160
161
162
        """Whether the task has a test set"""
        pass

163
    def training_docs(self) -> Iterable:
164
165
166
167
168
169
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

170
    def validation_docs(self) -> Iterable:
171
172
173
174
175
176
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

177
    def test_docs(self) -> Iterable:
178
179
180
181
182
183
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

184
    def fewshot_docs(self) -> Iterable:
185
186
187
188
189
190
191
192
193
        """
        :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
194
195
196
197
198
            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."
                )
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
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

Baber's avatar
Baber committed
219
    def fewshot_examples(self, k, 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
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
Baber committed
231
    def doc_to_text(self, doc: dict) -> str:
232
233
234
        pass

    @abc.abstractmethod
Baber's avatar
nit  
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
Baber committed
239
    def doc_to_image(self, doc: dict):
240
241
        raise NotImplementedError

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

Baber's avatar
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
nit  
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
nit  
Baber committed
257
        system_instruction: str | None = None,
258
259
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
nit  
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,
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
nit  
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
Baber's avatar
Baber committed
383
    def process_results(self, doc: dict, results: list):
384
385
386
387
388
389
390
391
392
393
394
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
        the metric for that one document

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param results:
            The results of the requests created in construct_requests.
        """
        pass

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
nit  
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
nit  
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
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
            # 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
480
            )
481
482

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

Baber's avatar
nit  
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"""
lintangsutawika's avatar
lintangsutawika committed
487
488
        if hasattr(self, "_filters"):
            for f in self._filters:
489
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
490
491
492
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
493

baberabb's avatar
baberabb committed
494
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
495
        """Returns the config as a dictionary."""
496
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
497
        # (num_fewshot)
498
        return self.config.to_dict()
499

Baber Abbasi's avatar
Baber Abbasi committed
500
501
502
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 key is None:
            raise ValueError("Key must be provided.")

        if update:
            current_value = getattr(self._config, key, {})
            if not isinstance(current_value, dict):
                raise TypeError(
                    f"Expected a dict for key '{key}', got {type(current_value).__name__} instead."
                )
            current_value.update(value)
        else:
            setattr(self._config, key, value)

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

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
Baber's avatar
Baber committed
522
523
524
525
526
        # 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)
        #     }
527
528
        self._config.metric_list = [MetricConfig(name=metric_name)]
        self._config.process_results = lambda *args: {"bypass": 0}
Baber Abbasi's avatar
Baber Abbasi committed
529

Baber's avatar
nit  
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
nit  
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
nit  
Baber committed
550
        limit: int | None = None,
551
        world_size: int = 1,
Baber's avatar
nit  
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
nit  
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
nit  
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
nit  
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())
662
663
        self.multiple_input = 0
        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
nit  
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
nit  
Baber committed
697
                choice_has_whitespace = choice[0].isspace()
698
                delimiter_has_whitespace = (
Baber's avatar
nit  
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's avatar
nit  
Baber committed
712
    def download(self, dataset_kwargs: dict[str, Any] | None = None, **kwargs) -> None:
713
714
715
716
717
        self.config.dataset_kwargs, self.config.metadata = (
            self.config.dataset_kwargs or {},
            self.config.metadata or {},
        )
        if isinstance(df := self.config.custom_dataset, Callable):
Baber Abbasi's avatar
Baber Abbasi committed
718
719
720
721
            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."
            )
722
            self.dataset = df(**(self.config.dataset_kwargs | self.config.metadata))
Baber Abbasi's avatar
Baber Abbasi committed
723
724
        else:
            self.dataset = datasets.load_dataset(
725
726
727
                path=self.config.dataset_path,
                name=self.config.dataset_name,
                **self.config.dataset_kwargs,
Baber Abbasi's avatar
Baber Abbasi committed
728
            )
729

baberabb's avatar
baberabb committed
730
    def has_training_docs(self) -> bool:
Baber's avatar
nit  
Baber committed
731
        return self.config.training_split is not None
732

baberabb's avatar
baberabb committed
733
    def has_validation_docs(self) -> bool:
Baber's avatar
nit  
Baber committed
734
        return self.config.validation_split is not None
735

baberabb's avatar
baberabb committed
736
    def has_test_docs(self) -> bool:
Baber's avatar
nit  
Baber committed
737
        return self.config.test_split is not None
738

Baber's avatar
nit  
Baber committed
739
    def training_docs(self) -> datasets.Dataset | None:
740
        if self.has_training_docs():
741
742
743
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
744
                )
745
            return self.dataset[self.config.training_split]
746

Baber's avatar
nit  
Baber committed
747
    def validation_docs(self) -> datasets.Dataset | None:
748
        if self.has_validation_docs():
749
750
751
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
752
                )
753
            return self.dataset[self.config.validation_split]
754

Baber's avatar
nit  
Baber committed
755
    def test_docs(self) -> datasets.Dataset | None:
756
        if self.has_test_docs():
757
758
759
            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]
760

761
    def fewshot_docs(self):
Baber's avatar
Baber committed
762
763
764
765
766
767
        docs = self.config.fewshot_cfg.get_docs(self.dataset)

        if docs is not None:
            return docs

        # Fallback to parent implementation
Baber's avatar
nit  
Baber committed
768
769
770
771
772
773
774
775
776
777
        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
778
779

        return super().fewshot_docs()
780

KonradSzafer's avatar
KonradSzafer committed
781
782
    @staticmethod
    def append_target_question(
Baber's avatar
nit  
Baber committed
783
        labeled_examples: list[dict[str, str]],
KonradSzafer's avatar
KonradSzafer committed
784
785
        question: str,
        fewshot_as_multiturn: bool = False,
Baber's avatar
nit  
Baber committed
786
        gen_prefix: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
787
788
789
790
791
792
793
794
    ) -> 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":
795
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
796
797
            # if last message is user, append to it to avoid two user messages in a row
            else:
798
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
799
800
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
801
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
802
803
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
804

lintangsutawika's avatar
lintangsutawika committed
805
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
806
807
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
808
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
809
        num_fewshot: int,
Baber's avatar
nit  
Baber committed
810
        system_instruction: str | None = None,
KonradSzafer's avatar
KonradSzafer committed
811
812
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
Baber's avatar
nit  
Baber committed
813
814
815
        chat_template: Callable | None = None,
        gen_prefix: str | None = None,
    ) -> str | list[str] | None:
lintangsutawika's avatar
lintangsutawika committed
816
817
818
819
820
821
822
        """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
823
824
825
826
827
828
        :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.
829
830
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
831
832
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
833
834
835
        :returns: str
            The fewshot context.
        """
Baber's avatar
nit  
Baber committed
836
        labeled_examples = [] if apply_chat_template else ""
KonradSzafer's avatar
KonradSzafer committed
837
838

        # get task description
839
840
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
841

KonradSzafer's avatar
KonradSzafer committed
842
843
844
845
846
847
848
849
850
        # 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
851
        else:
KonradSzafer's avatar
KonradSzafer committed
852
853
854
855
856
857
858
859
860
861
862
863
864
            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
865
866
867
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
868
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
869
870
871
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
872
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
873
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
874
                )
lintangsutawika's avatar
lintangsutawika committed
875
876

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
877
878
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
879
                # TODO: append prefill?
880
881
                if not labeled_examples:
                    return ""
882
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
883
884
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
885
886
887
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
888
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
889
890
891
892
893
894
895
                )
            # 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
896
897
898
899
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
900
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
901
902
903
904
905
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber's avatar
nit  
Baber committed
906
                            add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
907
908
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
909
910
911
912
913
914
                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
915
916
917
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
918
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
919
920
921
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
922
923
924
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
925
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
926
927
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
928
929
            return chat_template(
                labeled_examples,
Baber's avatar
nit  
Baber committed
930
                add_generation_prompt=not gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
931
            )
932
        else:
Baber Abbasi's avatar
Baber Abbasi committed
933
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
934
935
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
936
937
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
938
939
            if self.multiple_input:
                return labeled_examples
940
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
941
                return labeled_examples + example + prefix
942
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
943
                return [labeled_examples + ex + prefix for ex in example]
944
945
946
            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
947
                    return labeled_examples + choices[example] + prefix
948
                else:
Baber Abbasi's avatar
Baber Abbasi committed
949
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
950

Baber's avatar
nit  
Baber committed
951
    def apply_filters(self) -> list[Instance] | None:
Baber Abbasi's avatar
Baber Abbasi committed
952
        """Iterates over FilterEnsembles and applies them to instances"""
953
954
        if hasattr(self, "_filters"):
            for f in self._filters:
955
                f.ensemble.apply(self._instances)
956
957
958
959
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

960
    def should_decontaminate(self):
961
        return self.config.should_decontaminate
962

Baber Abbasi's avatar
Baber Abbasi committed
963
    def doc_to_decontamination_query(self, doc: dict):
964
        if self.config.should_decontaminate:
965
966
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
967
            else:
968
969
970
971
972
973
974
975
976
977
978
                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
                        )
                    )
979

980
    def _process_doc(self, doc: dict) -> dict:
981
982
983
984
985
986
987
988
989
990
        """
        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
nit  
Baber committed
991
    def doc_to_text(self, doc: dict, doc_to_text: int | str | Callable | None = None):
Baber's avatar
Baber committed
992
993
994
        # if self.prompt is not None:
        #     doc_to_text = self.prompt
        if doc_to_text is not None:
Yu Shi Jie's avatar
Yu Shi Jie committed
995
            doc_to_text = doc_to_text
996
        else:
997
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
998

999
        if isinstance(doc_to_text, int):
1000
            return doc_to_text
1001
        elif isinstance(doc_to_text, str):
1002
            if doc_to_text in self.features:
1003
                # if self.config.doc_to_choice is not None:
1004
1005
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1006
1007
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1008
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
nit  
Baber committed
1009
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1010
1011
1012
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1013
        elif callable(doc_to_text):
1014
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1015
        # Used when applying a Promptsource template
Baber's avatar
Baber committed
1016
1017
1018
1019
1020
1021
1022
        # 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
1023
        else:
1024
            print(type(doc_to_text))
1025
            raise TypeError
1026

Baber's avatar
nit  
Baber committed
1027
    def doc_to_target(self, doc: dict, doc_to_target=None) -> int | str | list[int]:
Baber's avatar
Baber committed
1028
1029
1030
        # 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
1031
            doc_to_target = doc_to_target
1032
        else:
1033
            doc_to_target = self.config.doc_to_target
1034

1035
        if isinstance(doc_to_target, int):
1036
            return doc_to_target
1037
        elif isinstance(doc_to_target, str):
1038
            if doc_to_target in self.features:
1039
                # if self.config.doc_to_choice is not None:
1040
1041
1042
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1043
            else:
lintangsutawika's avatar
lintangsutawika committed
1044
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
nit  
Baber committed
1045
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1046
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1047
1048
1049
1050
1051
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1052
1053
1054
1055
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1056
1057
                else:
                    return target_string
1058
        elif isinstance(doc_to_target, list):
1059
            return doc_to_target
1060
        elif callable(doc_to_target):
1061
            return doc_to_target(doc)
Baber's avatar
Baber committed
1062
1063
1064
1065
1066
1067
1068
1069
        # # 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
1070
1071
        else:
            raise TypeError
1072

Baber's avatar
Baber committed
1073
    def doc_to_choice(
Baber's avatar
Baber committed
1074
1075
        self,
        doc: dict,
Baber's avatar
nit  
Baber committed
1076
1077
        doc_to_choice: str | list | dict | Callable[..., list[str]] | None = None,
    ) -> list[str]:
Baber's avatar
Baber committed
1078
1079
1080
        # 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
1081
            doc_to_choice = doc_to_choice
1082
        elif self.config.doc_to_choice is None:
1083
            eval_logger.error("doc_to_choice was called but not set in config")
Baber's avatar
Baber committed
1084
            doc_to_choice = None
1085
        else:
1086
            doc_to_choice = self.config.doc_to_choice
1087

1088
        if isinstance(doc_to_choice, str):
1089
1090
1091
1092
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1093
        elif isinstance(doc_to_choice, list):
1094
            return doc_to_choice
1095
        elif isinstance(doc_to_choice, dict):
1096
1097
1098
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
Baber's avatar
Baber committed
1099
1100
        # elif hasattr(doc_to_choice, "get_answer_choices_list"):
        #     return doc_to_choice.get_answer_choices_list(doc)
1101
1102
        else:
            raise TypeError
1103

Baber's avatar
nit  
Baber committed
1104
    def doc_to_image(self, doc: dict, doc_to_image=None) -> int | str | list | None:
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
        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
nit  
Baber committed
1127
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> int | str | list | None:
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
        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
nit  
Baber committed
1150
    def doc_to_prefix(self, doc: dict) -> str | None:
Baber Abbasi's avatar
Baber Abbasi committed
1151
1152
1153
1154
1155
1156
1157
        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
1158
1159
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
Baber's avatar
nit  
Baber committed
1160
    ) -> list[Instance] | Instance:
1161
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1162
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1163

1164
1165
        aux_arguments = None

1166
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1167
            arguments = (ctx, self.doc_to_target(doc))
1168
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1169
            arguments = (self.doc_to_target(doc),)
1170
        elif self.OUTPUT_TYPE == "multiple_choice":
1171
            choices = self.doc_to_choice(doc)
1172
            target_delimiter = self.config.target_delimiter
1173
1174
            if apply_chat_template:
                target_delimiter = ""
1175
1176
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1177
                # apply chat_template to choices if apply_chat_template
1178
                cont = self.doc_to_target(doc)
1179

1180
                arguments = [
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1191
                ]
1192
            else:
1193
                # Otherwise they are placed in the continuation
1194
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1195

1196
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
Baber's avatar
Baber committed
1197
            if "acc_mutual_info" in [m.metric_name for m in self.config._metric_list]:
1198
1199
1200
1201
1202
1203
                # 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.
1204
1205
1206
1207
                # 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
                ]
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222

                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)},
            }

1223
1224
1225
1226
1227
1228
1229
1230
        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)},
            }

1231
1232
1233
1234
1235
1236
1237
        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":
1238
            request_list = [
1239
1240
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1241
                    doc=doc,
1242
                    arguments=arg,
1243
                    idx=i,
1244
1245
                    **kwargs,
                )
1246
                for i, arg in enumerate(arguments)
1247
            ]
1248
1249

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1250

lintangsutawika's avatar
lintangsutawika committed
1251
        return Instance(
1252
1253
1254
1255
1256
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1257
        )
1258

Baber's avatar
Baber committed
1259
    def process_results(self, doc: dict, results: list) -> dict:
1260
1261
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1262

1263
        result_dict = {}
Baber's avatar
fixup  
Baber committed
1264
        use_metric = list(m.metric_name for m in self.config._metric_list)
1265
1266
1267
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1268
1269
1270
1271
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1272
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1273
            (loglikelihood,) = results
1274
1275
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1276
            return {
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
                **(
                    {"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
1292
            }
1293
        elif self.OUTPUT_TYPE == "multiple_choice":
1294
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1295

Baber's avatar
nit  
Baber committed
1296
            # retrieve choices in list[str] form, to compute choice lengths, etc.
1297
            choices = self.doc_to_choice(doc)
1298
1299
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1300
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1301
1302
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1303
1304
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1305
1306
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1307
                # and this stores our "regular" conditional loglikelihoods
1308
                lls = lls[: len(choices)]
Baber's avatar
Baber committed
1309
1310
            else:
                lls_unconditional = None
1311

1312
1313
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1314

1315
1316
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1317
            else:
1318
                gold = self.doc_to_target(doc)
1319
1320

            gold_index_error = False
1321
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1322
1323
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1324
1325
                    gold_index_error = True
            else:
1326
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1327
                    gold = gold if gold < len(choices) else -100
1328
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1329
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1330

Lintang Sutawika's avatar
Lintang Sutawika committed
1331
                if gold == -100:
1332
1333
1334
1335
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1336
                    f"Label index was not in within range of available choices,"
1337
1338
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1339

1340
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1341
1342
                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
nit  
Baber committed
1343
                exact_match = int(any(is_greedy[i] if i != -100 else 0 for i in gold))
lintangsutawika's avatar
lintangsutawika committed
1344
1345
1346
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1347
                # 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
1348
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1349

Lintang Sutawika's avatar
Lintang Sutawika committed
1350
1351
1352
1353
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1354
            result_dict = {
1355
                **({"acc": acc} if "acc" in use_metric else {}),
1356
1357
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1358
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1359
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1360
1361
1362
1363
1364
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1365
1366
            }

1367
            if "acc_mutual_info" in use_metric:
Baber's avatar
Baber committed
1368
1369
1370
                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
1371
1372
1373
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1374
1375
1376
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1377
        elif self.OUTPUT_TYPE == "generate_until":
1378
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1379
            result = results[0]
1380
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1381
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1382
                # it assumes that doc_to_target returns a number.
1383
1384
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
Baber's avatar
nit  
Baber committed
1385
            for metric in self._metric_fn_list:
Baber's avatar
Baber committed
1386
1387
1388
1389
1390
1391
1392
1393
                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])
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:
Baber's avatar
fixup  
Baber committed
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:
Baber's avatar
fixup  
Baber committed
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
nit  
Baber committed
1419
    def task_name(self) -> str | None:
Lintang Sutawika's avatar
Lintang Sutawika committed
1420
1421
        return getattr(self.config, "task", None)

1422
1423
1424
1425
1426
    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
1427
            f"num_samples={len(self.eval_docs)})"
1428
1429
        )

1430
1431

class MultipleChoiceTask(Task):
1432
    OUTPUT_TYPE = "loglikelihood"
1433

baberabb's avatar
baberabb committed
1434
    def doc_to_target(self, doc: dict) -> str:
1435
1436
        return " " + doc["choices"][doc["gold"]]

Baber's avatar
nit  
Baber committed
1437
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> list[Instance]:
1438
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1439
1440
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1441
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1442
                doc=doc,
1443
                arguments=(ctx, f" {choice}"),
1444
                idx=i,
1445
1446
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1447
1448
            for i, choice in enumerate(doc["choices"])
        ]
1449

Baber's avatar
nit  
Baber committed
1450
    def process_results(self, doc: dict, results: Iterable[tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1451
1452
1453
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
        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
1465
    def higher_is_better(self) -> dict:
1466
1467
1468
1469
1470
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1471
    def aggregation(self) -> dict:
1472
1473
1474
1475
1476
1477
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1478
class PerplexityTask(Task):
1479
1480
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1481
    def has_training_docs(self) -> bool:
1482
1483
        return False

Baber's avatar
nit  
Baber committed
1484
    def fewshot_examples(self, k: int, rnd) -> list:
1485
1486
1487
1488
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1489
1490
        return []

baberabb's avatar
baberabb committed
1491
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1492
1493
1494
1495
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1496
1497
1498

        return ""

baberabb's avatar
baberabb committed
1499
    def higher_is_better(self) -> dict:
1500
1501
1502
1503
1504
1505
1506
1507
1508
        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
1509
    def doc_to_text(self, doc) -> str:
1510
1511
1512
1513
1514
        return ""

    def doc_to_target(self, doc):
        return doc

Baber's avatar
nit  
Baber committed
1515
    def construct_requests(self, doc: dict, ctx: str | None, **kwargs):
1516
1517
        if bool(ctx):
            raise ValueError
1518

lintangsutawika's avatar
lintangsutawika committed
1519
1520
1521
1522
1523
1524
1525
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1526

Baber's avatar
nit  
Baber committed
1527
    def process_results(self, doc: dict, results: tuple[float]) -> dict:
1528
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1529
1530
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1531
1532
1533
1534
1535
1536
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1537
    def aggregation(self) -> dict:
1538
1539
1540
1541
1542
1543
1544
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1545
    def count_bytes(cls, doc) -> int:
1546
1547
1548
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1549
    def count_words(cls, doc) -> int:
1550
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1551
        return len(re.split(r"\s+", doc))