task.py 77.5 KB
Newer Older
1
import abc
2
import ast
lintangsutawika's avatar
lintangsutawika committed
3
import logging
4
import random
5
6
import re
from collections.abc import Callable
7
from copy import deepcopy
Baber's avatar
Baber committed
8
from dataclasses import asdict, dataclass, field
Baber's avatar
Baber committed
9
from functools import cached_property
10
from inspect import getsource
11
12
13
14
15
16
17
18
19
20
21
22
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
23
24
25

import datasets
import numpy as np
26
from tqdm import tqdm
Baber's avatar
Baber committed
27
from typing_extensions import deprecated
28
29

from lm_eval import utils
30
from lm_eval.api import samplers
31
32
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
33
from lm_eval.api.registry import (
34
35
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
36
    get_aggregation,
37
    get_metric,
38
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
39
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
40
)
41
from lm_eval.caching.cache import load_from_cache, save_to_cache
42
43
44
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt

45

46
47
48
49
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
50
    "generate_until",
51
52
]

Lintang Sutawika's avatar
Lintang Sutawika committed
53
eval_logger = logging.getLogger(__name__)
54

lintangsutawika's avatar
lintangsutawika committed
55

Baber's avatar
Baber committed
56
57
58
59
60
61
62
63
64
65
@dataclass
class MetricConfig:
    """Encapsulates information about a single metric."""

    name: str
    fn: Optional[Callable] = None
    kwargs: Optional[dict] = None
    aggregation_fn: Optional[Callable] = None
    higher_is_better: bool = True
    hf_evaluate: bool = False
66
    is_elementwise: bool = True
Baber's avatar
Baber committed
67
68

    @cached_property
Baber's avatar
Baber committed
69
    def metric_name(self) -> str:
Baber's avatar
Baber committed
70
71
72
73
74
75
76
77
78
79
80
81
82
83
        return self.name

    @cached_property
    def aggregation(self) -> Callable:
        if self.aggregation_fn is None:
            return get_aggregation(self.name)
        return self.aggregation_fn

    @cached_property
    def _higher_is_better(self) -> bool:
        if self.higher_is_better is None:
            return is_higher_better(self.name)
        return self.higher_is_better

Baber's avatar
Baber committed
84
85
86
87
88
89
    def calculate_metric(self, *args, **kwargs) -> Any:
        """Calculates the metric using the provided function and arguments."""
        if self.fn is None:
            raise ValueError(f"Metric function for {self.name} is not defined.")
        return self.fn(*args, **{**self.kwargs, **kwargs})

Baber's avatar
Baber committed
90
91
92
93
94
95
    def compute_aggregation(self, values: List[Any]) -> Any:
        """Computes the aggregation of the metric values."""
        if self.aggregation_fn is None:
            raise ValueError(f"Aggregation function for {self.name} is not defined.")
        return self.aggregation_fn(values)

Baber's avatar
Baber committed
96

97
98
99
100
101
102
103
104
105
@dataclass
class RepeatConfig:
    """Encapsulates information about a single repeat."""

    repeats: int = 1
    metric_fn: Optional[Callable] = None
    kwargs: Optional[dict] = None


Baber's avatar
Baber committed
106
107
108
109
110
111
112
113
114
@dataclass
class FilterConfig:
    """Encapsulates information about a single filter."""

    name: str
    fn: Optional[Callable] = None
    kwargs: Optional[dict] = None


Baber's avatar
Baber committed
115
116
117
118
@dataclass
class FewshotConfig:
    sampler: str
    samples: list[dict]
119
    process_docs: Optional[Callable] = None
Baber's avatar
Baber committed
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
    fewshot_indices: Optional[list[int]] = None


@dataclass
class TemplateConfig:
    """Encapsulates information about a template."""

    template: str
    doc_to_text: Union[str, Callable[[dict], str]]
    doc_to_choice: Union[str, list, Callable[[dict], list]]
    doc_to_target: Union[int, Callable[[dict], int]]
    description: str
    context_prefix: str
    prefix_delimiter: str
    context_delimiter: str
    answer_suffix: str
    target_delimiter: str
    choice_format: Optional[str]
    choice_delimiter: Optional[str]
    fewshot_delimiter: str
    metric_list: Optional[Union[list[str], list[MetricConfig]]] = field(
        default_factory=lambda: ["acc", "acc_norm"]
    )


@dataclass
class MCQTemplateConfig:
    """Encapsulates information about a template.
    Would return a sample with the following format:
    Question: <doc_to_text(doc)>
    A. <doc_to_choice(doc)[0]>
    B. <doc_to_choice(doc)[1]>
    C. <doc_to_choice(doc)[2]>
    D. <doc_to_choice(doc)[3]>
    Answer:` doc_to_choice(doc)` for each choice.
    """

    doc_to_text: Union[str, Callable[[dict], str]]
    doc_to_choice: Union[str, list, Callable[[dict], list]]
    doc_to_target: Union[int, Callable[[dict], int]]
    template = "mcq"
    context_prefix: str = "Question:"
    prefix_delimiter: str = " "
    context_delimiter: str = "\n"
    answer_suffix: str = "Answer:"
    target_delimiter: str = "\n"
    choice_format: Optional[str] = "letters"
    choice_delimiter: Optional[str] = "\n"
    fewshot_delimiter: str = "\n\n"
    metric_list: Optional[list[MetricConfig]] = field(default_factory=lambda: ["acc"])


@dataclass
class ClozeTemplateConfig:
    """Encapsulates information about a template.
    Would return a sample with the following format:
    Question:  <doc_to_text(doc)>
    Answer:` <doc_to_target(doc)>`
    """

    doc_to_text: Union[str, Callable[[dict], str]]
    doc_to_choice: Union[str, list, Callable[[dict], list]]
    doc_to_target: Union[int, Callable[[dict], int]]
    template: str = "cloze"
    description: str = ""
    context_prefix: str = "Question:"
    prefix_delimiter: str = " "
    context_delimiter: str = "\n"
    answer_suffix: str = "Answer:"
    target_delimiter: str = " "
    choice_format: Optional[str] = None
    choice_delimiter: Optional[str] = None
    fewshot_delimiter: str = "\n\n"
    metric_list: Optional[list[MetricConfig]] = field(
        default_factory=lambda: ["acc", "acc_norm"]
    )
Baber's avatar
Baber committed
196
197


Baber's avatar
Baber committed
198
199
200
201
202
203
204
205
206
207
@dataclass
class DatasetConfig:
    """Encapsulates information about a dataset."""

    dataset_path: Optional[str] = None
    dataset_name: Optional[str] = None
    dataset_kwargs: Optional[dict] = None
    custom_dataset: Optional[Callable] = None


208
209
@dataclass
class TaskConfig(dict):
210
    # task naming/registry
211
212
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
213
    tag: Optional[Union[str, list]] = None
214
215
216
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
217
    custom_dataset: Optional[Callable] = None
218
219
220
221
222
223
    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
224
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
225
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
226
    )
227
228
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
229
230
231
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
Baber's avatar
Baber committed
232
233
    doc_to_image: Union[Callable, str, None] = None
    doc_to_audio: Union[Callable, str, None] = None
Hojin Lee's avatar
Hojin Lee committed
234
    unsafe_code: bool = False
235
236
237
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
238
    description: str = ""
239
240
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
241
    fewshot_config: Optional[dict] = None
242
    # runtime configuration options
243
    num_fewshot: Optional[int] = None
244
    # scoring options
245
246
247
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
248
    repeats: int = 1
249
    filter_list: Optional[Union[str, list]] = None
250
    should_decontaminate: bool = False
251
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
252
    gen_prefix: Optional[str] = None
253
254
255
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
Baber's avatar
Baber committed
256
257
    _metric_list = None
    _filter_list = None
258

Ethan Smith's avatar
Ethan Smith committed
259
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
260
        if self.generation_kwargs is not None:
261
            if self.output_type != "generate_until":
262
                eval_logger.warning(
263
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
264
265
266
267
268
269
270
271
                )

            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
272
273
274
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
275
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
276
        else:
277
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
278
279
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
280
281
282
283
284
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
285
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
286
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
287
                }
Baber Abbasi's avatar
Baber Abbasi committed
288
289
290
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
291

Baber's avatar
Baber committed
292
293
294
295
296
297
298
299
300
301
302
        if self.metric_list is not None:
            for metric_config in self.metric_list:
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )

    def get_metrics(self) -> list["MetricConfig"]:
        metrics = []
        if self.metric_list is None:
            _metric_list = DEFAULT_METRIC_REGISTRY[self.output_type]
Baber's avatar
Baber committed
303
304
305
            eval_logger.info(
                f"No metrics defined in config, using default metrics for {self.output_type}={_metric_list}"
            )
Baber's avatar
Baber committed
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
            metrics.extend(
                MetricConfig(
                    name=metric_name,
                    fn=get_metric(metric_name),
                    aggregation_fn=get_metric_aggregation(metric_name),
                    higher_is_better=is_higher_better(metric_name),
                )
                for metric_name in _metric_list
            )
        else:
            for metric_config in self.metric_list:
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
                metric_name = metric_config["metric"]
                _metric_fn_kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
                }
                _hf_evaluate_metric: bool = metric_config.get("hf_evaluate", False)
                _metric_fn = None
                _aggregation = None

                if self.process_results is not None:
                    # User will compute metrics inside `process_results()`
                    _metric_name = None
                    _metric_fn_kwargs = {}
                elif callable(metric_name):
                    # User passed a function object
                    _metric_name = metric_name.__name__
                    _metric_fn = metric_name.__call__
                else:
                    # Normal: look up by name
                    _metric_name = get_metric(metric_name, _hf_evaluate_metric)

                # ---------- 3. Decide how to aggregate examples ----------
                if "aggregation" in metric_config:
                    if isinstance(_agg_name := metric_config["aggregation"], str):
                        _aggregation = get_aggregation(_agg_name)
                    elif callable(_agg_name):  # noqa: E721
                        _aggregation = metric_config["aggregation"]
                else:
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
                    _aggregation = get_metric_aggregation(metric_name)
                    eval_logger.warning(
                        f"[Task: {self.task}] metric {metric_name} is defined, but aggregation is not. "
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[_aggregation]}"
                    )

                # ---------- 4. Determine “higher-is-better” semantics ----------
                if "higher_is_better" in metric_config:
                    _higher_is_better = metric_config["higher_is_better"]
                else:
                    eval_logger.warning(
                        f"[Task: {self.task}] metric {metric_name} is defined, but higher_is_better is not. "
                        f"using default "
                        f"higher_is_better={is_higher_better(metric_name)}"
                    )
                    _higher_is_better = is_higher_better(metric_name)

                metrics.append(
                    MetricConfig(
                        name=_metric_name,
                        fn=_metric_fn,
                        kwargs=_metric_fn_kwargs,
                        aggregation_fn=_aggregation,
                        higher_is_better=_higher_is_better,
                        hf_evaluate=_hf_evaluate_metric,
                    )
                )
        return metrics

Baber's avatar
Baber committed
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
    def get_filters(self):
        if self.filter_list is not None:
            _filter_list = []
            if isinstance(self.filter_list, dict):
                for filter_config in self.filter_list:
                    _filter_list.append(
                        build_filter_ensemble(
                            filter_name=filter_config["name"],
                            components=[
                                [
                                    {
                                        key: function[key]
                                        for key in function
                                        if key != "function"
                                    }
                                ]
                                for function in filter_config["filter"]
                            ],
                        )
                    )
        else:
            # 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."
            )
            _filter_list = [build_filter_ensemble("none", [["take_first", None]])]

        return _filter_list

411
412
413
    def __getitem__(self, item):
        return getattr(self, item)

414
415
416
    def __setitem__(self, item, value):
        return setattr(self, item, value)

417
    def to_dict(self, keep_callable: bool = False) -> dict:
418
419
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
420
        Used for dumping results alongside full task configuration
421

haileyschoelkopf's avatar
haileyschoelkopf committed
422
423
424
425
426
427
428
429
430
431
        :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)
432
433
434
435
436
437
438
439
440
441
            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
442
        return cfg_dict
443

444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
    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)

460
461
462
463
464
465
466
467
468
469
470

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)
    """

471
    VERSION: Optional[Union[int, str]] = None
472

473
474
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
475
    DATASET_PATH: Optional[str] = None
476
477

    # The name of a subset within `DATASET_PATH`.
478
    DATASET_NAME: Optional[str] = None
479

480
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
481

482
483
    def __init__(
        self,
484
485
486
487
        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
488
    ) -> None:
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
        """
        :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)
511
512
513
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
514

515
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
516

lintangsutawika's avatar
lintangsutawika committed
517
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
518
519
520
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
521

522
523
524
525
526
527
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
        """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.
        """
552
553
554
555
556
557
558
        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,
        )
559

560
    @property
561
    def config(self) -> TaskConfig:
562
563
564
        """Returns the TaskConfig associated with this class."""
        return self._config

565
    @abc.abstractmethod
Baber's avatar
Baber committed
566
    def has_training_docs(self) -> bool:
567
568
569
570
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
571
    def has_validation_docs(self) -> bool:
572
573
574
575
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
576
    def has_test_docs(self) -> bool:
577
578
579
        """Whether the task has a test set"""
        pass

580
    def training_docs(self) -> Iterable:
581
582
583
584
585
586
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

587
    def validation_docs(self) -> Iterable:
588
589
590
591
592
593
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

594
    def test_docs(self) -> Iterable:
595
596
597
598
599
600
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

601
    def fewshot_docs(self) -> Iterable:
602
603
604
605
606
607
608
609
610
        """
        :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
611
612
613
614
615
            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."
                )
616
617
            return self.test_docs()

618
    def _process_doc(self, doc: dict) -> dict:
619
620
621
622
623
624
625
626
627
        """
        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
628

629
    @property
630
    def instances(self) -> List[Instance]:
631
632
633
634
635
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

Baber's avatar
Baber committed
636
    def fewshot_examples(self, k, rnd) -> Iterable[dict]:
637
638
639
640
641
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

642
643
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
644
645
646
647
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
Baber committed
648
    def doc_to_text(self, doc) -> str:
649
650
651
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
652
    def doc_to_target(self, doc) -> Union[str, int]:
653
654
        pass

655
656
657
658
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

659
660
661
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber's avatar
Baber committed
662
    def doc_to_prefix(self, doc) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
663
664
        return ""

665
666
    def build_all_requests(
        self,
667
        *,
668
        limit: Union[int, None] = None,
669
        samples: Optional[List[int]] = None,
670
671
672
673
674
675
676
677
678
        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 = "",
679
    ) -> None:
680
        """Build a set of Instances for a task, and store them in task.instances"""
681
682
683
684

        # used with caching
        og_limit = limit

685
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
686
687
688
689
690
691
692
        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 ""
        )
693
        cache_key += f"-tokenizer{tokenizer_name}"
694

Baber Abbasi's avatar
Baber Abbasi committed
695
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
696
697
698
699
700
701
702
703
704
705
706
707
708

        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
709
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
710

711
        instances = []
712
713
714
715
716
717
718
719
720
721

        # 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(
722
723
724
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
725
726
727
728
729
730
731
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
732
        ):
733
            # sample fewshot context #TODO: need to offset doc_id by rank now!
734
            fewshot_ctx = self.fewshot_context(
735
                doc,
736
737
738
739
740
741
742
                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
743
                gen_prefix=self.doc_to_prefix(doc),
744
            )
745

746
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
747
748
749
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
750
                metadata=(self.config["task"], doc_id, self.config.repeats),
751
                apply_chat_template=apply_chat_template,
752
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
753
            )
754
755
756
757

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

758
759
760
761
762
763
764
765
766
767
768
769
770
            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
771

772
773
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
774

775
776
777
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
    @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
794
            The number of times each instance in a dataset is inferred on. Defaults to 1,
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
            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

Baber's avatar
Baber committed
812
    @deprecated("not used anymore")
813
814
815
816
817
818
819
820
    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

Baber's avatar
Baber committed
821
    @deprecated("not used anymore")
822
823
824
825
826
827
828
829
    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

830
831
832
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
833
    @classmethod
Baber's avatar
Baber committed
834
    def count_bytes(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
835
836
837
838
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
839
    def count_words(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
840
841
842
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

843
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
844
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
845
846
847
848
849
850
851
        """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
852
853
854
855
856
        :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.
857
858
859
        :returns: str
            The fewshot context.
        """
860
        if rnd is None:
861
862
863
864
865
866
            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
867

868
        description = description if description else ""
869
870

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
871
            labeled_examples = ""
872
        else:
lintangsutawika's avatar
lintangsutawika committed
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
            # 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
897
            )
898
899

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
900
        return description + labeled_examples + example
901

902
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
903
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
904
905
        if hasattr(self, "_filters"):
            for f in self._filters:
906
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
907
908
909
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
910

baberabb's avatar
baberabb committed
911
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
912
        """Returns the config as a dictionary."""
913
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
914
        # (num_fewshot)
915
        return self.config.to_dict()
916

Baber Abbasi's avatar
Baber Abbasi committed
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if key is None:
            raise ValueError("Key must be provided.")

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

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

        Parameters:
        - metric_name (str): The name of the custom metric to override. Should be registered in api.metrics.
        """
Baber's avatar
Baber committed
939
940
941
942
943
944
945
        # if not isinstance(self, ConfigurableTask):
        #     self.process_results = lambda x, y: {metric_name: get_metric(metric_name)}
        #     self.aggregation = lambda: {
        #         metric_name: get_metric_aggregation(metric_name)
        #     }
        setattr(self._config, "metric_list", [MetricConfig(name=metric_name)])
        setattr(self._config, "process_results", lambda *args: {"bypass": 0})
Baber Abbasi's avatar
Baber Abbasi committed
946

947
948
949
950
951
    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

952
    @property
Baber's avatar
Baber committed
953
    def eval_docs(self) -> Union[datasets.Dataset, Iterable[dict]]:
954
955
956
957
958
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
959
960
961
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
962
963

    def doc_iterator(
964
965
966
967
968
969
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
970
    ) -> Iterator[Tuple[int, Any]]:
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
        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),
            )
993
994
        return doc_iterator

995
996

class ConfigurableTask(Task):
997
    VERSION = "Yaml"
998
    OUTPUT_TYPE = None
999
    CONFIG = None
1000
1001

    def __init__(
1002
1003
1004
1005
1006
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Baber's avatar
Baber committed
1007
    ) -> None:
1008
        # Get pre-configured attributes
1009
        self._config = self.CONFIG
1010

1011
        # Use new configurations if there was no preconfiguration
1012
        if self.config is None:
1013
            self._config = TaskConfig(**config)
1014
1015
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
1016
            if config is not None:
1017
                self._config.__dict__.update(config)
1018

1019
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
1020
1021
1022
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
1023

1024
1025
1026
1027
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

1028
        if self.config.output_type is not None:
1029
1030
1031
1032
            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)}'"
                )
1033
            self.OUTPUT_TYPE = self.config.output_type
1034

1035
1036
1037
1038
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

1039
1040
1041
1042
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
1043
1044
1045
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

1046
1047
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
1048

1049
1050
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
1051

Baber's avatar
Baber committed
1052
        self.metric_list: list[MetricConfig] = self.config.get_metrics()
1053

1054
        self.download(self.config.dataset_kwargs)
1055
1056
1057
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
1058
1059
        self._filters = self.config.get_filters()

1060
1061
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
1062
            self.prompt = get_prompt(
1063
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
1064
            )
1065
1066
1067
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
1068
        if self.fewshot_docs() is not None:
1069
1070
1071
1072
            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
1073
1074
1075
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
            )
            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)}"
                )
1092

1093
        self.task_docs = self.eval_docs
1094

1095
        # Test One Doc
1096
        self.features = list(self.task_docs.features.keys())
1097
1098
        self.multiple_input = 0
        self.multiple_target = 0
1099
        test_doc = self.task_docs[0]
1100
        test_text = self.doc_to_text(test_doc)
1101
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
1102

1103
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1104
            test_choice = self.doc_to_choice(test_doc)
1105
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
1106
                eval_logger.error("doc_to_choice must return list")
1107
1108
            else:
                num_choice = len(test_choice)
1109

1110
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
1111
1112
1113
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
1114
                self.multiple_input = num_choice
1115
1116
        else:
            test_choice = None
1117

1118
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
1119
1120
1121
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
1122
            self.multiple_target = len(test_target)
1123
        else:
1124
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1125
                test_target = test_choice[test_target]
1126
            else:
lintangsutawika's avatar
lintangsutawika committed
1127
                test_target = str(test_target)
1128

1129
1130
1131
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1132
            check_choices = [test_target]
1133
1134
1135
1136
        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 = (
1137
1138
                    True
                    if self.config.target_delimiter.rstrip()
1139
                    != self.config.target_delimiter
1140
                    else False
1141
                )
1142

1143
                if delimiter_has_whitespace and choice_has_whitespace:
1144
1145
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1146
1147
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1148
                    eval_logger.debug(
1149
                        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'
1150
1151
                    )

Baber Abbasi's avatar
Baber Abbasi committed
1152
1153
1154
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
Baber Abbasi's avatar
Baber Abbasi committed
1155
1156
1157
1158
        from packaging.version import parse as vparse

        if dataset_kwargs and vparse(datasets.__version__) >= vparse("4.0.0"):
            dataset_kwargs.pop("trust_remote_code", None)
Baber Abbasi's avatar
Baber Abbasi committed
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
        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 {},
            )
1173

baberabb's avatar
baberabb committed
1174
    def has_training_docs(self) -> bool:
1175
        if self.config.training_split is not None:
1176
1177
1178
1179
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1180
    def has_validation_docs(self) -> bool:
1181
        if self.config.validation_split is not None:
1182
1183
1184
1185
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1186
    def has_test_docs(self) -> bool:
1187
        if self.config.test_split is not None:
1188
1189
1190
1191
            return True
        else:
            return False

Baber's avatar
Baber committed
1192
    def training_docs(self) -> Optional[datasets.Dataset]:
1193
        if self.has_training_docs():
1194
1195
1196
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1197
                )
1198
            return self.dataset[self.config.training_split]
1199

Baber's avatar
Baber committed
1200
    def validation_docs(self) -> Optional[datasets.Dataset]:
1201
        if self.has_validation_docs():
1202
1203
1204
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1205
                )
1206
            return self.dataset[self.config.validation_split]
1207

Baber's avatar
Baber committed
1208
    def test_docs(self) -> Optional[datasets.Dataset]:
1209
        if self.has_test_docs():
1210
1211
1212
            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]
1213

1214
    def fewshot_docs(self):
1215
        if self.config.fewshot_split is not None:
1216
1217
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1218
            return self.dataset[self.config.fewshot_split]
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
        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."
                )
1231
        else:
1232
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1233
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1234
                    f"[Task: {self.config.task}] "
1235
1236
1237
1238
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1239

KonradSzafer's avatar
KonradSzafer committed
1240
1241
1242
1243
1244
    @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
1245
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1246
1247
1248
1249
1250
1251
1252
1253
    ) -> 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":
1254
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1255
1256
            # if last message is user, append to it to avoid two user messages in a row
            else:
1257
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1258
1259
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1260
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1261
1262
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1263

lintangsutawika's avatar
lintangsutawika committed
1264
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1265
1266
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1267
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1268
1269
1270
1271
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1272
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1273
        gen_prefix: Optional[str] = None,
Baber's avatar
Baber committed
1274
    ) -> Union[str, List[str], None]:
lintangsutawika's avatar
lintangsutawika committed
1275
1276
1277
1278
1279
1280
1281
        """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
1282
1283
1284
1285
1286
1287
        :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.
1288
1289
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1290
1291
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1292
1293
1294
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1295
1296
1297
1298
1299
1300
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1301
1302
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1303

KonradSzafer's avatar
KonradSzafer committed
1304
1305
1306
1307
1308
1309
1310
1311
1312
        # 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
1313
        else:
KonradSzafer's avatar
KonradSzafer committed
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
            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
1327
1328
1329
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1330
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1331
1332
1333
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1334
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1335
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1336
                )
lintangsutawika's avatar
lintangsutawika committed
1337
1338

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1339
1340
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1341
                # TODO: append prefill?
1342
1343
                if not labeled_examples:
                    return ""
1344
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1345
1346
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1347
1348
1349
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1350
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1351
1352
1353
1354
1355
1356
1357
                )
            # 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
1358
1359
1360
1361
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1362
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1363
1364
1365
1366
1367
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1368
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1369
1370
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1371
1372
1373
1374
1375
1376
                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
1377
1378
1379
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1380
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1381
1382
1383
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1384
1385
1386
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1387
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1388
1389
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1390
1391
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1392
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1393
            )
1394
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1395
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1396
1397
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1398
1399
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1400
1401
            if self.multiple_input:
                return labeled_examples
1402
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1403
                return labeled_examples + example + prefix
1404
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1405
                return [labeled_examples + ex + prefix for ex in example]
1406
1407
1408
            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
1409
                    return labeled_examples + choices[example] + prefix
1410
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1411
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1412

Baber Abbasi's avatar
Baber Abbasi committed
1413
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1414
        """Iterates over FilterEnsembles and applies them to instances"""
1415
1416
        if hasattr(self, "_filters"):
            for f in self._filters:
1417
                f.apply(self._instances)
1418
1419
1420
1421
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1422
    def should_decontaminate(self):
1423
        return self.config.should_decontaminate
1424

Baber Abbasi's avatar
Baber Abbasi committed
1425
    def doc_to_decontamination_query(self, doc: dict):
1426
        if self.config.should_decontaminate:
1427
1428
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1429
            else:
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
                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
                        )
                    )
1441

1442
    def _process_doc(self, doc: dict) -> dict:
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
        """
        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
1453
    def doc_to_text(self, doc, doc_to_text=None):
1454
1455
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1456
1457
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1458
        else:
1459
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1460

1461
        if isinstance(doc_to_text, int):
1462
            return doc_to_text
1463
        elif isinstance(doc_to_text, str):
1464
            if doc_to_text in self.features:
1465
                # if self.config.doc_to_choice is not None:
1466
1467
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1468
1469
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1470
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
Baber committed
1471
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1472
1473
1474
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1475
        elif callable(doc_to_text):
1476
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1477
        # Used when applying a Promptsource template
1478
        elif hasattr(doc_to_text, "apply"):
1479
1480
1481
1482
1483
            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")
1484
                return self.config.fewshot_delimiter
1485
        else:
1486
            print(type(doc_to_text))
1487
            raise TypeError
1488

Yu Shi Jie's avatar
Yu Shi Jie committed
1489
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1490
1491
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1492
1493
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1494
        else:
1495
            doc_to_target = self.config.doc_to_target
1496

1497
        if isinstance(doc_to_target, int):
1498
            return doc_to_target
1499
        elif isinstance(doc_to_target, str):
1500
            if doc_to_target in self.features:
1501
                # if self.config.doc_to_choice is not None:
1502
1503
1504
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1505
            else:
lintangsutawika's avatar
lintangsutawika committed
1506
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
Baber committed
1507
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1508
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1509
1510
1511
1512
1513
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1514
1515
1516
1517
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1518
1519
                else:
                    return target_string
1520
        elif isinstance(doc_to_target, list):
1521
            return doc_to_target
1522
        elif callable(doc_to_target):
1523
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1524
        # Used when applying a Promptsource template
1525
        elif hasattr(doc_to_target, "apply"):
1526
            applied_prompt = doc_to_target.apply(doc)
1527
1528
1529
1530
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1531
                return self.config.fewshot_delimiter
1532
1533
        else:
            raise TypeError
1534

Yu Shi Jie's avatar
Yu Shi Jie committed
1535
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1536
1537
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1538
1539
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1540
        elif self.config.doc_to_choice is None:
1541
1542
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1543
            doc_to_choice = self.config.doc_to_choice
1544

1545
        if isinstance(doc_to_choice, str):
1546
1547
1548
1549
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1550
        elif isinstance(doc_to_choice, list):
1551
            return doc_to_choice
1552
        elif isinstance(doc_to_choice, dict):
1553
1554
1555
1556
1557
1558
1559
            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
1560

Baber's avatar
Baber committed
1561
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list, None]:
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

Baber's avatar
Baber committed
1584
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list, None]:
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber's avatar
Baber committed
1607
    def doc_to_prefix(self, doc) -> Optional[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1608
1609
1610
1611
1612
1613
1614
        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
1615
1616
1617
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1618
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1619
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1620

1621
1622
        aux_arguments = None

1623
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1624
            arguments = (ctx, self.doc_to_target(doc))
1625
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1626
            arguments = (self.doc_to_target(doc),)
1627
        elif self.OUTPUT_TYPE == "multiple_choice":
1628
            choices = self.doc_to_choice(doc)
1629
            target_delimiter = self.config.target_delimiter
1630
1631
            if apply_chat_template:
                target_delimiter = ""
1632
1633
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1634
                # apply chat_template to choices if apply_chat_template
1635
                cont = self.doc_to_target(doc)
1636

1637
                arguments = [
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1648
                ]
1649
            else:
1650
                # Otherwise they are placed in the continuation
1651
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1652

1653
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
Baber's avatar
Baber committed
1654
            if "acc_mutual_info" in [m.metric_name for m in self.metric_list]:
1655
1656
1657
1658
1659
1660
                # 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.
1661
1662
1663
1664
                # 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
                ]
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679

                arguments.extend(aux_arguments)

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

        multimodal_arg = {}
        if (
            self.config.doc_to_image
        ):  # TODO: ensure that non-multimodal tasks aren't getting visual args
            multimodal_arg = {
                **multimodal_arg,
                **{"visual": self.doc_to_image(doc)},
            }

1680
1681
1682
1683
1684
1685
1686
1687
        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)},
            }

1688
1689
1690
1691
1692
1693
1694
        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":
1695
            request_list = [
1696
1697
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1698
                    doc=doc,
1699
                    arguments=arg,
1700
                    idx=i,
1701
1702
                    **kwargs,
                )
1703
                for i, arg in enumerate(arguments)
1704
            ]
1705
1706

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1707

lintangsutawika's avatar
lintangsutawika committed
1708
        return Instance(
1709
1710
1711
1712
1713
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1714
        )
1715
1716

    def process_results(self, doc, results):
1717
1718
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1719

1720
        result_dict = {}
Baber's avatar
Baber committed
1721
        use_metric = list(m.metric_name for m in self.metric_list)
1722
1723
1724
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1725
1726
1727
1728
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1729
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1730
            (loglikelihood,) = results
1731
1732
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1733
            return {
1734
1735
1736
1737
1738
1739
1740
1741
1742
1743
1744
1745
1746
1747
1748
                **(
                    {"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
1749
            }
1750
        elif self.OUTPUT_TYPE == "multiple_choice":
1751
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1752

1753
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1754
            choices = self.doc_to_choice(doc)
1755
1756
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1757
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1758
1759
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1760
1761
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1762
1763
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1764
                # and this stores our "regular" conditional loglikelihoods
1765
                lls = lls[: len(choices)]
1766

1767
1768
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1769

1770
1771
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1772
            else:
1773
                gold = self.doc_to_target(doc)
1774
1775

            gold_index_error = False
1776
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1777
1778
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1779
1780
                    gold_index_error = True
            else:
1781
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1782
                    gold = gold if gold < len(choices) else -100
1783
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1784
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1785

Lintang Sutawika's avatar
Lintang Sutawika committed
1786
                if gold == -100:
1787
1788
1789
1790
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1791
                    f"Label index was not in within range of available choices,"
1792
1793
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1794

1795
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1796
1797
                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
1798
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1799
1800
1801
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1802
                # 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
1803
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1804

Lintang Sutawika's avatar
Lintang Sutawika committed
1805
1806
1807
1808
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1809
            result_dict = {
1810
                **({"acc": acc} if "acc" in use_metric else {}),
1811
1812
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1813
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1814
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1815
1816
1817
1818
1819
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1820
1821
            }

1822
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1823
1824
1825
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1826
1827
1828
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1829
        elif self.OUTPUT_TYPE == "generate_until":
1830
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1831
            result = results[0]
1832
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1833
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1834
                # it assumes that doc_to_target returns a number.
1835
1836
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1837
1838
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1839
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1840
1841
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
Baber's avatar
Baber committed
1842
                "bypass" in use_metric or isinstance(result, list)
1843
            ):
Chris's avatar
Chris committed
1844
1845
                # cast gold to the same type as result
                gold = type(result)(gold)
1846

Baber's avatar
Baber committed
1847
            for metric in self.metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
1848
1849
1850
1851
1852
                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
1853
1854
1855
1856
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
Baber's avatar
Baber committed
1857
                    if metric.name == "exact_match":
1858
                        result = [result for _ in range(len(gold))]
Baber's avatar
Baber committed
1859
                        scores = metric.fn(
1860
1861
                            references=gold,
                            predictions=result,
Baber's avatar
Baber committed
1862
                            **metric.kwargs,
1863
1864
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1865
                    else:
1866
1867
                        for gold_option in gold:
                            try:
Baber's avatar
Baber committed
1868
                                result_score = metric.fn(
1869
1870
                                    references=[gold_option],
                                    predictions=[result],
Baber's avatar
Baber committed
1871
                                    **metric.kwargs,
1872
1873
1874
1875
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
Baber's avatar
Baber committed
1876
                                result_score = metric.fn([gold_option, result])
1877
1878
1879
1880
1881
1882
1883
1884
                            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
1885
                else:
1886
                    try:
Baber's avatar
Baber committed
1887
                        result_score = metric.fn(
1888
1889
                            references=[gold],
                            predictions=[result],
Baber's avatar
Baber committed
1890
                            **metric.kwargs,
1891
                        )
1892
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
Baber's avatar
Baber committed
1893
                        result_score = metric.fn([gold, result])
1894
1895
1896
1897
1898
1899
1900
                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
1901
        else:
lintangsutawika's avatar
lintangsutawika committed
1902
1903
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1904
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1905
            )
1906
1907
1908

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1909
    def aggregation(self) -> dict:
Baber's avatar
Baber committed
1910
        return {k.name: k.aggregation_fn for k in self.metric_list}
1911

Baber Abbasi's avatar
Baber Abbasi committed
1912
    def higher_is_better(self) -> dict:
Baber's avatar
Baber committed
1913
        return {k.name: k.higher_is_better for k in self.metric_list}
1914

Baber Abbasi's avatar
Baber Abbasi committed
1915
1916
1917
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1918
    @property
Baber's avatar
Baber committed
1919
    def task_name(self) -> Optional[str]:
Lintang Sutawika's avatar
Lintang Sutawika committed
1920
1921
        return getattr(self.config, "task", None)

1922
1923
1924
1925
1926
    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
1927
            f"num_samples={len(self.eval_docs)})"
1928
1929
        )

1930
1931

class MultipleChoiceTask(Task):
1932
    OUTPUT_TYPE = "loglikelihood"
1933

baberabb's avatar
baberabb committed
1934
    def doc_to_target(self, doc: dict) -> str:
1935
1936
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1937
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1938
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1939
1940
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1941
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1942
                doc=doc,
1943
                arguments=(ctx, " {}".format(choice)),
1944
                idx=i,
1945
1946
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1947
1948
            for i, choice in enumerate(doc["choices"])
        ]
1949

1950
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1951
1952
1953
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
        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
1965
    def higher_is_better(self) -> dict:
1966
1967
1968
1969
1970
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1971
    def aggregation(self) -> dict:
1972
1973
1974
1975
1976
1977
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1978
class PerplexityTask(Task):
1979
1980
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1981
    def has_training_docs(self) -> bool:
1982
1983
        return False

baberabb's avatar
baberabb committed
1984
    def fewshot_examples(self, k: int, rnd) -> List:
1985
1986
1987
1988
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1989
1990
        return []

baberabb's avatar
baberabb committed
1991
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1992
1993
1994
1995
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1996
1997
1998

        return ""

baberabb's avatar
baberabb committed
1999
    def higher_is_better(self) -> dict:
2000
2001
2002
2003
2004
2005
2006
2007
2008
        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
2009
    def doc_to_text(self, doc) -> str:
2010
2011
2012
2013
2014
        return ""

    def doc_to_target(self, doc):
        return doc

2015
2016
2017
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
2018

lintangsutawika's avatar
lintangsutawika committed
2019
2020
2021
2022
2023
2024
2025
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
2026

2027
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
2028
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
2029
2030
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
2031
2032
2033
2034
2035
2036
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
2037
    def aggregation(self) -> dict:
2038
2039
2040
2041
2042
2043
2044
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
2045
    def count_bytes(cls, doc) -> int:
2046
2047
2048
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
2049
    def count_words(cls, doc) -> int:
2050
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
2051
        return len(re.split(r"\s+", doc))