task.py 76.1 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
import re
Baber's avatar
test  
Baber committed
6
from collections import defaultdict
7
from collections.abc import Callable
8
from copy import deepcopy
9
from dataclasses import asdict, dataclass
10
from inspect import getsource
Baber's avatar
test  
Baber committed
11
12
from itertools import groupby
from operator import attrgetter
13
14
15
16
17
18
19
20
21
22
23
24
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
25
26
27

import datasets
import numpy as np
28
from tqdm import tqdm
29
30

from lm_eval import utils
31
from lm_eval.api import samplers
32
33
from lm_eval.api.instance import Instance, OutputType
from lm_eval.api.metrics import bits_per_byte, mean, weighted_perplexity
lintangsutawika's avatar
lintangsutawika committed
34
from lm_eval.api.registry import (
35
36
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
37
    get_aggregation,
38
    get_metric,
39
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
40
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
41
)
Baber's avatar
Baber committed
42
from lm_eval.api.schemas import MetricResult
43
from lm_eval.caching.cache import load_from_cache, save_to_cache
44
45
46
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

47

48
49
50
51
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
52
    "generate_until",
53
54
]

Lintang Sutawika's avatar
Lintang Sutawika committed
55
eval_logger = logging.getLogger(__name__)
56

lintangsutawika's avatar
lintangsutawika committed
57

58
59
@dataclass
class TaskConfig(dict):
60
    # task naming/registry
61
62
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
63
    tag: Optional[Union[str, list]] = None
64
65
66
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
67
    custom_dataset: Optional[Callable] = None
68
69
70
71
72
73
    dataset_path: Optional[str] = None
    dataset_name: Optional[str] = None
    dataset_kwargs: Optional[dict] = None
    training_split: Optional[str] = None
    validation_split: Optional[str] = None
    test_split: Optional[str] = None
74
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
75
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
76
    )
77
78
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
79
80
81
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
82
    doc_to_image: Union[Callable, str] = None
83
    doc_to_audio: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
84
    unsafe_code: bool = False
85
86
87
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
88
    description: str = ""
89
90
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
91
    fewshot_config: Optional[dict] = None
92
    # runtime configuration options
93
    num_fewshot: Optional[int] = None
94
    # scoring options
95
96
97
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
98
    repeats: int = 1
99
    filter_list: Optional[Union[str, list]] = None
100
    should_decontaminate: bool = False
101
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
102
    gen_prefix: Optional[str] = None
103
    repeat_agg: Optional[str] = None
104
105
106
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
107

Ethan Smith's avatar
Ethan Smith committed
108
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
109
        if self.generation_kwargs is not None:
110
            if self.output_type != "generate_until":
111
                eval_logger.warning(
112
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
113
114
115
116
117
118
119
120
                )

            if "temperature" in self.generation_kwargs:
                self.generation_kwargs["temperature"] = float(
                    self.generation_kwargs["temperature"]
                )

            if "until" not in self.generation_kwargs:
Baber Abbasi's avatar
Baber Abbasi committed
121
122
123
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
124
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
125
        else:
126
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
127
128
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
129
130
131
132
133
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
134
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
135
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
136
                }
Baber Abbasi's avatar
Baber Abbasi committed
137
138
139
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
140

141
142
143
    def __getitem__(self, item):
        return getattr(self, item)

144
145
146
    def __setitem__(self, item, value):
        return setattr(self, item, value)

147
    def to_dict(self, keep_callable: bool = False) -> dict:
148
149
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
150
        Used for dumping results alongside full task configuration
151

haileyschoelkopf's avatar
haileyschoelkopf committed
152
153
154
155
156
157
158
159
160
161
        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
            if v is None:
                cfg_dict.pop(k)
162
163
164
165
166
167
168
169
170
171
            elif k == "metric_list":
                for metric_dict in v:
                    for metric_key, metric_value in metric_dict.items():
                        if callable(metric_value):
                            metric_dict[metric_key] = self.serialize_function(
                                metric_value, keep_callable=keep_callable
                            )
                cfg_dict[k] = v
            elif callable(v):
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
haileyschoelkopf's avatar
haileyschoelkopf committed
172
        return cfg_dict
173

174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
    def serialize_function(
        self, value: Union[Callable, str], keep_callable=False
    ) -> Union[Callable, str]:
        """Serializes a given function or string.

        If 'keep_callable' is True, the original callable is returned.
        Otherwise, attempts to return the source code of the callable using 'getsource'.
        """
        if keep_callable:
            return value
        else:
            try:
                return getsource(value)
            except (TypeError, OSError):
                return str(value)

190
191
192
193
194
195
196
197
198
199
200

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)
    """

201
    VERSION: Optional[Union[int, str]] = None
202

203
204
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
205
    DATASET_PATH: Optional[str] = None
206
207

    # The name of a subset within `DATASET_PATH`.
208
    DATASET_NAME: Optional[str] = None
209

210
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
211

212
213
    def __init__(
        self,
214
215
216
217
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode: Optional[datasets.DownloadMode] = None,
        config: Optional[Mapping] = None,  # Union[dict, TaskConfig]
Ethan Smith's avatar
Ethan Smith committed
218
    ) -> None:
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
        """
        :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)
241
242
243
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
244

245
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
246

lintangsutawika's avatar
lintangsutawika committed
247
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
248
249
250
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
251

252
253
254
255
256
257
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
        """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.
        """
282
283
284
285
286
287
288
        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,
        )
289

290
    @property
291
    def config(self) -> TaskConfig:
292
293
294
        """Returns the TaskConfig associated with this class."""
        return self._config

295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
    def has_validation_docs(self):
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
    def has_test_docs(self):
        """Whether the task has a test set"""
        pass

310
    def training_docs(self) -> Iterable:
311
312
313
314
315
316
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

317
    def validation_docs(self) -> Iterable:
318
319
320
321
322
323
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

324
    def test_docs(self) -> Iterable:
325
326
327
328
329
330
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

331
    def fewshot_docs(self) -> Iterable:
332
333
334
335
336
337
338
339
340
        """
        :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
341
342
343
344
345
            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."
                )
346
347
            return self.test_docs()

348
    def _process_doc(self, doc: dict) -> dict:
349
350
351
352
353
354
355
356
357
        """
        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
358

359
    @property
360
    def instances(self) -> List[Instance]:
361
362
363
364
365
366
367
368
369
370
371
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

    def fewshot_examples(self, k, rnd):
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

372
373
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
374
375
376
377
378
379
380
381
382
383
384
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
    def doc_to_text(self, doc):
        pass

    @abc.abstractmethod
    def doc_to_target(self, doc):
        pass

385
386
387
388
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

389
390
391
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
392
393
394
    def doc_to_prefix(self, doc):
        return ""

395
396
    def build_all_requests(
        self,
397
        *,
398
        limit: Union[int, None] = None,
399
        samples: Optional[List[int]] = None,
400
401
402
403
404
405
406
407
408
        rank: int = 0,
        world_size: int = 1,
        cache_requests: bool = False,
        rewrite_requests_cache: bool = False,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
        chat_template: Optional[Callable] = None,
        tokenizer_name: str = "",
409
    ) -> None:
410
        """Build a set of Instances for a task, and store them in task.instances"""
411
412
413
414

        # used with caching
        og_limit = limit

415
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
416
417
418
419
420
421
422
        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 ""
        )
423
        cache_key += f"-tokenizer{tokenizer_name}"
424

Baber Abbasi's avatar
Baber Abbasi committed
425
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
426
427
428
429
430
431
432
433
434
435
436
437
438

        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
439
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
440

441
        instances = []
442
443
444
445
446
447
448
449
450
451

        # 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(
452
453
454
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
455
456
457
458
459
460
461
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
462
        ):
463
            # sample fewshot context #TODO: need to offset doc_id by rank now!
464
            fewshot_ctx = self.fewshot_context(
465
                doc,
466
467
468
469
470
471
472
                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
473
                gen_prefix=self.doc_to_prefix(doc),
474
            )
475

476
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
477
478
479
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
480
                metadata=(self.config["task"], doc_id, self.config.repeats),
481
                apply_chat_template=apply_chat_template,
482
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
483
            )
484
485
486
487

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

488
489
490
491
492
493
494
495
496
497
498
499
500
            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
501

502
503
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
504

505
506
507
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
    @abc.abstractmethod
    def construct_requests(self, doc, ctx, **kwargs):
        """Uses RequestFactory to construct Requests and returns an iterable of
        Requests which will be sent to the LM.

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
            The context string, generated by fewshot_context. This includes the natural
            language description, as well as the few shot examples, and the question
            part of the document for `doc`.
        :param doc_idx: int
            The index of a document within `self.test_docs()` or `self.validation_docs()`,
            whichever is the main split used.
        :param repeats: int
        TODO: update this docstring
lintangsutawika's avatar
lintangsutawika committed
524
            The number of times each instance in a dataset is inferred on. Defaults to 1,
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
    def process_results(self, doc, results):
        """Take a single document and the LM results and evaluates, returning a
        dict where keys are the names of submetrics and values are the values of
        the metric for that one document

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

    @abc.abstractmethod
    def aggregation(self):
        """
        :returns: {str: [metric_score] -> float}
            A dictionary where keys are the names of submetrics and values are
            functions that aggregate a list of metric scores
        """
        pass

    @abc.abstractmethod
    def higher_is_better(self):
        """
        :returns: {str: bool}
            A dictionary where keys are the names of submetrics and values are
            whether a higher value of the submetric is better
        """
        pass

560
561
562
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
563
564
565
566
567
568
569
570
571
572
    @classmethod
    def count_bytes(cls, doc):
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
    def count_words(cls, doc):
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

573
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
574
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
575
576
577
578
579
580
581
        """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
582
583
584
585
586
        :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.
587
588
589
        :returns: str
            The fewshot context.
        """
590
        if rnd is None:
591
592
593
594
595
596
            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
597

598
        description = description if description else ""
599
600

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
601
            labeled_examples = ""
602
        else:
lintangsutawika's avatar
lintangsutawika committed
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
            # 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
627
            )
628
629

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
630
        return description + labeled_examples + example
631

632
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
633
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
634
635
        if hasattr(self, "_filters"):
            for f in self._filters:
636
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
637
638
639
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
640

baberabb's avatar
baberabb committed
641
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
642
        """Returns the config as a dictionary."""
643
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
644
        # (num_fewshot)
645
        return self.config.to_dict()
646

Baber Abbasi's avatar
Baber Abbasi committed
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if key is None:
            raise ValueError("Key must be provided.")

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

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

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
        (
            self._metric_fn_list,
            self._aggregation_list,
            self._metric_fn_kwargs,
            self._higher_is_better,
        ) = ({}, {}, {}, {})
        self._metric_fn_list[metric_name] = get_metric(metric_name)
        self._aggregation_list[metric_name] = get_metric_aggregation(metric_name)
        self._higher_is_better[metric_name] = is_higher_better(metric_name)
        self._metric_fn_kwargs[metric_name] = {}
        if not isinstance(self, ConfigurableTask):
            self.process_results = lambda x, y: {metric_name: get_metric(metric_name)}
            self.aggregation = lambda: {
                metric_name: get_metric_aggregation(metric_name)
            }
        setattr(self._config, "metric_list", [{"metric": metric_name}])
        setattr(self._config, "process_results", None)

687
688
689
690
691
    def set_fewshot_seed(self, seed: Optional[int] = None) -> None:
        self.fewshot_rnd = random.Random(seed)
        if hasattr(self, "sampler"):
            self.sampler.rnd = self.fewshot_rnd

692
693
694
695
696
697
698
    @property
    def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
699
700
701
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
702
703

    def doc_iterator(
704
705
706
707
708
709
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
710
    ) -> Iterator[Tuple[int, Any]]:
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
        if samples:
            n = len(self.eval_docs)
            assert all([e < n for e in samples]), (
                f"Elements of --samples should be in the interval [0,k-1] where k is the number of total examples. In this case, k={n}."
            )
            eval_logger.info(
                f"{self.config.task}: Evaluating on {len(samples)} examples"
            )
            doc_iterator = utils.create_iterator(
                enumerate(x for i, x in enumerate(self.eval_docs) if i in samples),
                rank=int(rank),
                limit=None,  # limit does not matter here since we are selecting samples directly
                world_size=int(world_size),
            )
        else:
            limit = int(limit) if limit else None
            doc_iterator = utils.create_iterator(
                enumerate(self.eval_docs),
                rank=int(rank),
                limit=limit,
                world_size=int(world_size),
            )
733
734
        return doc_iterator

735
736

class ConfigurableTask(Task):
737
    VERSION = "Yaml"
738
    OUTPUT_TYPE = None
739
    CONFIG = None
740
741

    def __init__(
742
743
744
745
746
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
747
    ) -> None:  # TODO no super() call here
748
        # Get pre-configured attributes
749
        self._config = self.CONFIG
750

751
        # Use new configurations if there was no preconfiguration
752
        if self.config is None:
753
            self._config = TaskConfig(**config)
754
755
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
756
            if config is not None:
757
                self._config.__dict__.update(config)
758

759
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
760
761
762
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
763

764
765
766
767
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

768
        if self.config.output_type is not None:
769
770
771
772
            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)}'"
                )
773
            self.OUTPUT_TYPE = self.config.output_type
774

775
776
777
778
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

779
780
781
782
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
783
784
785
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

786
787
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
788

789
790
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
791

792
793
794
795
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
796

797
        if self.config.metric_list is None:
798
            # TODO: handle this in TaskConfig.__post_init__ ?
799
800
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

801
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
802
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
803
                self._metric_fn_kwargs[metric_name] = {}
804
805
806
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
807
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
808
        else:
809
            for metric_config in self.config.metric_list:
810
811
812
813
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
814
815
816
817
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
818
819
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
820
                }
Chris's avatar
Chris committed
821
822
823
824
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
825

826
                if self.config.process_results is not None:
827
828
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
829
830
831
832
833
834
                elif callable(metric_name):
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                    self._metric_fn_list[metric_name] = metric_fn
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
Chris's avatar
Chris committed
835
836
837
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
838
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
839

840
                if "aggregation" in metric_config:
841
                    agg_name = metric_config["aggregation"]
842
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
843
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
844
                    elif callable(agg_name):  # noqa: E721
845
846
847
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
848
                else:
849
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
850
                    metric_agg = get_metric_aggregation(metric_name)
851
                    eval_logger.warning(
852
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
853
854
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
855
                    )
856
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
857

858
859
860
861
862
863
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
864
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
865
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
866
                        f"higher_is_better={is_higher_better(metric_name)}"
867
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
868
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
869

870
        self.download(self.config.dataset_kwargs)
871
872
873
        self._training_docs = None
        self._fewshot_docs = None

874
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
875
            self._filters = []
876
            for filter_config in self.config.filter_list:
877
878
879
880
881
882
883
884
885
                filter_name = filter_config["name"]
                filter_functions = filter_config["filter"]
                components = []
                for function in filter_functions:
                    kwargs = {
                        key: function[key] for key in function if key != "function"
                    }
                    components.append([function["function"], kwargs])
                filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
886
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
887
        else:
Baber Abbasi's avatar
Baber Abbasi committed
888
889
890
891
            # TODO: handle repeats in a more general way rather than just discarding
            eval_logger.debug(
                "No custom filters defined. Using default 'take_first' filter for handling repeats."
            )
892
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
893

894
895
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
896
            self.prompt = get_prompt(
897
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
898
            )
899
900
901
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
902
        if self.fewshot_docs() is not None:
903
904
905
906
            self.fewshot_rnd = (
                random.Random()
            )  # setting with no seed, to be overridden at a later time
            config_sampler: Union[str, Callable] = (
haileyschoelkopf's avatar
haileyschoelkopf committed
907
908
909
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
            )
            if isinstance(config_sampler, str):
                self.sampler = samplers.get_sampler(config_sampler)(
                    list(self.fewshot_docs()), self, rnd=self.fewshot_rnd
                )
            elif callable(config_sampler) and issubclass(
                config_sampler, samplers.ContextSampler
            ):
                self.sampler = config_sampler(
                    docs=list(self.fewshot_docs()), task=self, rnd=self.fewshot_rnd
                )
            else:
                raise TypeError(
                    f"fewshot_config.sampler should be a string or callable of ContextSampler type, "
                    f"not {type(config_sampler)}"
                )
926

927
        self.task_docs = self.eval_docs
928

929
        # Test One Doc
930
        self.features = list(self.task_docs.features.keys())
931
932
        self.multiple_input = 0
        self.multiple_target = 0
933
        test_doc = self.task_docs[0]
934
        test_text = self.doc_to_text(test_doc)
935
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
936

937
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
938
            test_choice = self.doc_to_choice(test_doc)
939
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
940
                eval_logger.error("doc_to_choice must return list")
941
942
            else:
                num_choice = len(test_choice)
943

944
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
945
946
947
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
948
                self.multiple_input = num_choice
949
950
        else:
            test_choice = None
951

952
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
953
954
955
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
956
            self.multiple_target = len(test_target)
957
        else:
958
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
959
                test_target = test_choice[test_target]
960
            else:
lintangsutawika's avatar
lintangsutawika committed
961
                test_target = str(test_target)
962

963
964
965
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
966
            check_choices = [test_target]
967
968
969
970
        if self.config.doc_to_choice is not None:
            for choice in check_choices:
                choice_has_whitespace = True if choice[0].isspace() else False
                delimiter_has_whitespace = (
971
972
                    True
                    if self.config.target_delimiter.rstrip()
973
                    != self.config.target_delimiter
974
                    else False
975
                )
976

977
                if delimiter_has_whitespace and choice_has_whitespace:
978
979
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
980
981
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
982
                    eval_logger.debug(
983
                        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'
984
985
                    )

Baber Abbasi's avatar
Baber Abbasi committed
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
1001
1002
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
        if isinstance(self.config.custom_dataset, Callable):
            eval_logger.warning(
                f"{self.config.task}: Custom kwargs can be passed to `--metadata` in console (as json string) or to the TaskManager."
                + "\nFor example --metadata='{\"max_seq_lengths\":[4096, 8192]}'. For details see task Readme."
            )
            self.dataset = self.config.custom_dataset(
                **(self.config.metadata or {}), **(self.config.dataset_kwargs or {})
            )
        else:
            self.dataset = datasets.load_dataset(
                path=self.DATASET_PATH,
                name=self.DATASET_NAME,
                **dataset_kwargs if dataset_kwargs is not None else {},
            )
1003

baberabb's avatar
baberabb committed
1004
    def has_training_docs(self) -> bool:
1005
        if self.config.training_split is not None:
1006
1007
1008
1009
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1010
    def has_validation_docs(self) -> bool:
1011
        if self.config.validation_split is not None:
1012
1013
1014
1015
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1016
    def has_test_docs(self) -> bool:
1017
        if self.config.test_split is not None:
1018
1019
1020
1021
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1022
    def training_docs(self) -> datasets.Dataset:
1023
        if self.has_training_docs():
1024
1025
1026
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1027
                )
1028
            return self.dataset[self.config.training_split]
1029

baberabb's avatar
baberabb committed
1030
    def validation_docs(self) -> datasets.Dataset:
1031
        if self.has_validation_docs():
1032
1033
1034
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1035
                )
1036
            return self.dataset[self.config.validation_split]
1037

baberabb's avatar
baberabb committed
1038
    def test_docs(self) -> datasets.Dataset:
1039
        if self.has_test_docs():
1040
1041
1042
            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]
1043

1044
    def fewshot_docs(self):
1045
        if self.config.fewshot_split is not None:
1046
1047
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1048
            return self.dataset[self.config.fewshot_split]
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
        elif (
            self.config.fewshot_config is not None
            and self.config.fewshot_config.get("samples", None) is not None
        ):
            if isinstance(self.config.fewshot_config["samples"], list):
                return self.config.fewshot_config["samples"]
            elif callable(self.config.fewshot_config["samples"]):
                return self.config.fewshot_config["samples"]()
            else:
                raise Exception(
                    "`fewshot_config['samples']` was incorrectly defined in the configuration. It should be either a list of samples as a dict, or function returning this list."
                )
1061
        else:
1062
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1063
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1064
                    f"[Task: {self.config.task}] "
1065
1066
1067
1068
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1069

KonradSzafer's avatar
KonradSzafer committed
1070
1071
1072
1073
1074
    @staticmethod
    def append_target_question(
        labeled_examples: List[Dict[str, str]],
        question: str,
        fewshot_as_multiturn: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
1075
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1076
1077
1078
1079
1080
1081
1082
1083
    ) -> 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":
1084
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1085
1086
            # if last message is user, append to it to avoid two user messages in a row
            else:
1087
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1088
1089
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1090
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1091
1092
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1093

lintangsutawika's avatar
lintangsutawika committed
1094
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1095
1096
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1097
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1098
1099
1100
1101
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1102
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1103
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1104
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1105
1106
1107
1108
1109
1110
1111
        """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
1112
1113
1114
1115
1116
1117
        :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.
1118
1119
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1120
1121
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1122
1123
1124
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1125
1126
1127
1128
1129
1130
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1131
1132
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1133

KonradSzafer's avatar
KonradSzafer committed
1134
1135
1136
1137
1138
1139
1140
1141
1142
        # 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
1143
        else:
KonradSzafer's avatar
KonradSzafer committed
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
            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
1157
1158
1159
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1160
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1161
1162
1163
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1164
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1165
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1166
                )
lintangsutawika's avatar
lintangsutawika committed
1167
1168

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

Baber Abbasi's avatar
Baber Abbasi committed
1243
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1244
        """Iterates over FilterEnsembles and applies them to instances"""
1245
1246
        if hasattr(self, "_filters"):
            for f in self._filters:
1247
                f.apply(self._instances)
1248
1249
1250
1251
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1252
    def should_decontaminate(self):
1253
        return self.config.should_decontaminate
1254

Baber Abbasi's avatar
Baber Abbasi committed
1255
    def doc_to_decontamination_query(self, doc: dict):
1256
        if self.config.should_decontaminate:
1257
1258
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1259
            else:
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
                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
                        )
                    )
1271

1272
    def _process_doc(self, doc: dict) -> dict:
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

        :return: dict
            The processed version of the specified `doc`.
        """
        return doc

Yu Shi Jie's avatar
Yu Shi Jie committed
1283
    def doc_to_text(self, doc, doc_to_text=None):
1284
1285
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1286
1287
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1288
        else:
1289
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1290

1291
        if isinstance(doc_to_text, int):
1292
            return doc_to_text
1293
        elif isinstance(doc_to_text, str):
1294
            if doc_to_text in self.features:
1295
                # if self.config.doc_to_choice is not None:
1296
1297
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1298
1299
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1300
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1301
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1302
1303
1304
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1305
        elif callable(doc_to_text):
1306
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1307
        # Used when applying a Promptsource template
1308
        elif hasattr(doc_to_text, "apply"):
1309
1310
1311
1312
1313
            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")
1314
                return self.config.fewshot_delimiter
1315
        else:
1316
            print(type(doc_to_text))
1317
            raise TypeError
1318

Yu Shi Jie's avatar
Yu Shi Jie committed
1319
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1320
1321
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1322
1323
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1324
        else:
1325
            doc_to_target = self.config.doc_to_target
1326

1327
        if isinstance(doc_to_target, int):
1328
            return doc_to_target
1329
        elif isinstance(doc_to_target, str):
1330
            if doc_to_target in self.features:
1331
                # if self.config.doc_to_choice is not None:
1332
1333
1334
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1335
            else:
lintangsutawika's avatar
lintangsutawika committed
1336
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1337
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1338
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1339
1340
1341
1342
1343
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1344
1345
1346
1347
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1348
1349
                else:
                    return target_string
1350
        elif isinstance(doc_to_target, list):
1351
            return doc_to_target
1352
        elif callable(doc_to_target):
1353
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1354
        # Used when applying a Promptsource template
1355
        elif hasattr(doc_to_target, "apply"):
1356
            applied_prompt = doc_to_target.apply(doc)
1357
1358
1359
1360
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1361
                return self.config.fewshot_delimiter
1362
1363
        else:
            raise TypeError
1364

Yu Shi Jie's avatar
Yu Shi Jie committed
1365
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1366
1367
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1368
1369
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1370
        elif self.config.doc_to_choice is None:
1371
1372
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1373
            doc_to_choice = self.config.doc_to_choice
1374

1375
        if isinstance(doc_to_choice, str):
1376
1377
1378
1379
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1380
        elif isinstance(doc_to_choice, list):
1381
            return doc_to_choice
1382
        elif isinstance(doc_to_choice, dict):
1383
1384
1385
1386
1387
1388
1389
            return list(doc_to_choice.values())
        elif callable(doc_to_choice):
            return doc_to_choice(doc)
        elif hasattr(doc_to_choice, "get_answer_choices_list"):
            return doc_to_choice.get_answer_choices_list(doc)
        else:
            raise TypeError
1390

1391
1392
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list]:
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list]:
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber Abbasi's avatar
Baber Abbasi committed
1437
1438
1439
1440
1441
1442
1443
1444
    def doc_to_prefix(self, doc):
        if (gen_prefix := self.config.gen_prefix) is not None:
            if gen_prefix in self.features:
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1445
1446
1447
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1448
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1449
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1450

1451
1452
        aux_arguments = None

1453
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1454
            arguments = (ctx, self.doc_to_target(doc))
1455
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1456
            arguments = (self.doc_to_target(doc),)
1457
        elif self.OUTPUT_TYPE == "multiple_choice":
1458
            choices = self.doc_to_choice(doc)
1459
            target_delimiter = self.config.target_delimiter
1460
1461
            if apply_chat_template:
                target_delimiter = ""
1462
1463
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1464
                # apply chat_template to choices if apply_chat_template
1465
                cont = self.doc_to_target(doc)
1466

1467
                arguments = [
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1478
                ]
1479
            else:
1480
                # Otherwise they are placed in the continuation
1481
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1482

1483
1484
1485
1486
1487
1488
1489
1490
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # log(P(choice|ctx) / P(choice)) = log(P(choice|ctx)) - log(P(choice))
                # in other words normalizing by subtracting the unconditional logprob of each choice.
1491
1492
1493
1494
                # 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
                ]
1495
1496
1497
1498
1499
1500

                arguments.extend(aux_arguments)

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

Baber's avatar
Baber committed
1501
1502
1503
1504
1505
1506
1507
        else:
            raise ValueError(
                f"Unsupported OUTPUT_TYPE: '{self.OUTPUT_TYPE}'. "
                f"Expected one of: 'loglikelihood', 'loglikelihood_rolling', "
                f"'multiple_choice', 'generate_until'"
            )

1508
1509
1510
1511
1512
1513
1514
1515
1516
        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)},
            }

1517
1518
1519
1520
1521
1522
1523
1524
        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)},
            }

1525
1526
1527
1528
1529
1530
1531
        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":
1532
            request_list = [
1533
1534
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1535
                    doc=doc,
Baber's avatar
Baber committed
1536
1537
                    arguments=arg,
                    # arguments=LoglikelihoodInput(context=arg[0], continuation=arg[1]),
1538
                    idx=i,
1539
1540
                    **kwargs,
                )
1541
                for i, arg in enumerate(arguments)
1542
            ]
1543
1544

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1545

lintangsutawika's avatar
lintangsutawika committed
1546
        return Instance(
1547
1548
            request_type=self.OUTPUT_TYPE,
            doc=doc,
Baber's avatar
Baber committed
1549
1550
1551
            arguments=arguments,
            # if self.OUTPUT_TYPE in ["loglikelihood", "loglikelihood_rolling"]
            # else GenerateInput(*arguments),
1552
1553
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1554
        )
1555
1556

    def process_results(self, doc, results):
1557
1558
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1559

1560
        result_dict = {}
1561
        use_metric = list(self._metric_fn_list.keys())
1562
1563
1564
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1565
1566
1567
1568
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1569
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1570
            (loglikelihood,) = results
1571
1572
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1573
            return {
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
                **(
                    {"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
1589
            }
1590
        elif self.OUTPUT_TYPE == "multiple_choice":
1591
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1592

1593
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1594
            choices = self.doc_to_choice(doc)
1595
1596
            completion_len = np.array([float(len(i)) for i in choices])

1597
1598
            if (
                2 * len(choices) == len(lls)
1599
                and "acc_mutual_info" in self._metric_fn_list.keys()
1600
1601
1602
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1603
1604
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1605
1606
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1607
                # and this stores our "regular" conditional loglikelihoods
1608
                lls = lls[: len(choices)]
1609

1610
1611
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1612

1613
1614
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1615
            else:
1616
                gold = self.doc_to_target(doc)
1617
1618

            gold_index_error = False
1619
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1620
1621
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1622
1623
                    gold_index_error = True
            else:
1624
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1625
                    gold = gold if gold < len(choices) else -100
1626
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1627
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1628

Lintang Sutawika's avatar
Lintang Sutawika committed
1629
                if gold == -100:
1630
1631
1632
1633
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1634
                    f"Label index was not in within range of available choices,"
1635
1636
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1637

1638
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1639
1640
                acc = 1.0 if pred in gold else 0.0
                acc_norm = 1.0 if pred_norm in gold else 0.0
Lintang Sutawika's avatar
Lintang Sutawika committed
1641
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1642
1643
1644
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1645
                # 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
1646
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1647

Lintang Sutawika's avatar
Lintang Sutawika committed
1648
1649
1650
1651
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1652
            result_dict = {
1653
                **({"acc": acc} if "acc" in use_metric else {}),
1654
1655
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1656
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1657
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1658
1659
1660
1661
1662
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1663
1664
            }

1665
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1666
1667
1668
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1669
1670
1671
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1672
        elif self.OUTPUT_TYPE == "generate_until":
1673
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1674
            result = results[0]
1675
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1676
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1677
                # it assumes that doc_to_target returns a number.
1678
1679
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1680
1681
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1682
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1683
1684
1685
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
1686
            ):
Chris's avatar
Chris committed
1687
1688
                # cast gold to the same type as result
                gold = type(result)(gold)
1689

lintangsutawika's avatar
lintangsutawika committed
1690
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1691
1692
1693
1694
1695
                if self.multiple_target:
                    # in the case where we have multiple targets,
                    # return true if any are true
                    # TODO: this may break for multipLe_target, non zero-or-1 metrics
                    scores = []
haileyschoelkopf's avatar
haileyschoelkopf committed
1696
1697
1698
1699
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1700
1701
1702
1703
1704
1705
1706
1707
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1708
                    else:
1709
1710
1711
1712
1713
1714
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1730
                else:
1731
                    try:
1732
                        result_score = self._metric_fn_list[metric](
1733
1734
                            references=[gold],
                            predictions=[result],
1735
                            **self._metric_fn_kwargs[metric],
1736
                        )
1737
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1738
                        result_score = self._metric_fn_list[metric]([gold, result])
1739
1740
1741
1742
1743
1744
1745
                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
1746
        else:
lintangsutawika's avatar
lintangsutawika committed
1747
1748
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1749
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1750
            )
1751
1752
1753

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1754
    def aggregation(self) -> dict:
1755
1756
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1757
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1758
        return self._higher_is_better
1759

Baber Abbasi's avatar
Baber Abbasi committed
1760
1761
1762
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1763
1764
1765
1766
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1767
1768
1769
1770
1771
    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
1772
            f"num_samples={len(self.eval_docs)})"
1773
1774
        )

1775
    def calculate_metrics(
Baber's avatar
TODO!  
Baber committed
1776
        self,
1777
        instances_by_doc_id=None,
Baber's avatar
TODO!  
Baber committed
1778
1779
1780
1781
1782
        filter_keys=None,
        samples=None,
        rank=1,
        limit=None,
        world_size=1,
Baber's avatar
test  
Baber committed
1783
    ) -> list[MetricResult]:
1784
1785
1786
1787
1788
1789
1790
1791
1792
1793
1794
1795
1796
        """Calculate metrics for all datapoints in the task.

        Args:
            instances_by_doc_id (dict): Dictionary mapping doc_ids to lists of instances.
            filter_key (str): The filter key to use for filtered responses.
            samples (dict, optional): Dictionary of sample indices to evaluate.
            rank (int): The process rank.
            limit (int, optional): Limit on number of examples to evaluate.
            world_size (int): Total number of processes.

        Returns:
            list: A list of metrics calculated for each document.
        """
1797
1798
1799
        if not self._instances:
            return

Baber's avatar
TODO!  
Baber committed
1800
        if filter_keys is None:
1801
1802
1803
1804
1805
            filter_keys = (
                [x.name for x in self._filters]
                if hasattr(self, "_filters")
                else ["none"]
            )
Baber's avatar
TODO!  
Baber committed
1806
1807
        if isinstance(filter_keys, str):
            filter_keys = [filter_keys]
1808
1809
1810
1811
        if not instances_by_doc_id:
            instances_by_doc_id = defaultdict(list)
            for instance in self.instances:
                instances_by_doc_id[instance.doc_id].append(instance)
Baber's avatar
test  
Baber committed
1812
1813
        # all_metrics = collections.defaultdict(list)
        all_metrics = []
Baber's avatar
TODO!  
Baber committed
1814
1815
1816
1817
1818
1819
1820
        for filter_key in filter_keys:
            doc_iterator = self.doc_iterator(
                rank=rank,
                limit=limit,
                world_size=world_size,
                # samples=indices,
            )
1821

Baber's avatar
TODO!  
Baber committed
1822
1823
1824
            for doc_id, doc in doc_iterator:
                # doc_id_true = indices[doc_id] if indices else doc_id
                requests = instances_by_doc_id[doc_id]
Baber's avatar
Baber committed
1825
1826
1827
1828
                if len(requests) > 1:
                    # if one doc has multiple instances then calculate metric together
                    metrics = self.process_results(
                        doc, [req.filtered_resps[filter_key] for req in requests]
Baber's avatar
TODO!  
Baber committed
1829
                    )
Baber's avatar
Baber committed
1830
1831
1832
1833
1834
1835
1836
1837
1838
1839
                else:
                    metrics = [
                        self.process_results(doc, response)
                        for req in requests
                        for response in (
                            req.filtered_resps[filter_key]
                            if isinstance(req.filtered_resps[filter_key], list)
                            else [req.filtered_resps[filter_key]]
                        )
                    ]
Baber's avatar
test  
Baber committed
1840
                all_metrics.append(
1841
1842
                    MetricResult(scores=metrics, doc_id=doc_id, filter_key=filter_key)
                )
1843
1844
1845

        return all_metrics

Baber's avatar
test  
Baber committed
1846
1847
1848
1849
1850
1851
1852
1853
1854
1855
1856
1857
1858
    @staticmethod
    def compute_agg_metrics(self, metric_results: list[MetricResult]):
        y_sorted = sorted(metric_results, key=attrgetter("filter_key", "metric_name"))

        groups = {
            key: list(
                map(list, zip(*((d[it.metric_name] for d in it.scores) for it in g)))
            )
            for key, g in groupby(y_sorted, key=attrgetter("filter_key", "metric_name"))
        }

        return groups

1859
1860

class MultipleChoiceTask(Task):
1861
    OUTPUT_TYPE = "loglikelihood"
1862

baberabb's avatar
baberabb committed
1863
    def doc_to_target(self, doc: dict) -> str:
1864
1865
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1866
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1867
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1868
1869
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1870
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1871
                doc=doc,
1872
                arguments=(ctx, " {}".format(choice)),
1873
                idx=i,
1874
1875
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1876
1877
            for i, choice in enumerate(doc["choices"])
        ]
1878

1879
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1880
1881
1882
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1883
1884
1885
1886
1887
1888
1889
1890
1891
1892
1893
        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
1894
    def higher_is_better(self) -> dict:
1895
1896
1897
1898
1899
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1900
    def aggregation(self) -> dict:
1901
1902
1903
1904
1905
1906
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1907
class PerplexityTask(Task):
Baber's avatar
Baber committed
1908
    OUTPUT_TYPE: OutputType = "loglikelihood_rolling"
1909

baberabb's avatar
baberabb committed
1910
    def has_training_docs(self) -> bool:
1911
1912
        return False

baberabb's avatar
baberabb committed
1913
    def fewshot_examples(self, k: int, rnd) -> List:
1914
1915
1916
1917
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1918
1919
        return []

baberabb's avatar
baberabb committed
1920
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1921
1922
1923
1924
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1925
1926
1927

        return ""

baberabb's avatar
baberabb committed
1928
    def higher_is_better(self) -> dict:
1929
1930
1931
1932
1933
1934
1935
1936
1937
        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
1938
    def doc_to_text(self, doc) -> str:
1939
1940
1941
1942
1943
        return ""

    def doc_to_target(self, doc):
        return doc

1944
1945
1946
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1947

lintangsutawika's avatar
lintangsutawika committed
1948
1949
1950
1951
1952
1953
1954
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1955

1956
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1957
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1958
1959
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1960
1961
1962
1963
1964
1965
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1966
    def aggregation(self) -> dict:
1967
1968
1969
1970
1971
1972
1973
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1974
    def count_bytes(cls, doc) -> int:
1975
1976
1977
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1978
    def count_words(cls, doc) -> int:
1979
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1980
        return len(re.split(r"\s+", doc))