task.py 76.7 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
from typing import (
Baber's avatar
Baber committed
12
    TYPE_CHECKING,
13
14
15
16
17
18
19
20
21
22
23
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
24
25
26

import datasets
import numpy as np
27
from tqdm import tqdm
Baber's avatar
Baber committed
28
from typing_extensions import deprecated
29
30

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

46

47
48
49
50
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
51
    "generate_until",
52
53
]

Baber's avatar
Baber committed
54
55
56
57
if TYPE_CHECKING:
    from lm_eval.api.filter import FilterEnsemble


Lintang Sutawika's avatar
Lintang Sutawika committed
58
eval_logger = logging.getLogger(__name__)
59

lintangsutawika's avatar
lintangsutawika committed
60

Baber's avatar
Baber committed
61
62
63
64
65
66
67
68
69
70
@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
71
    is_elementwise: bool = True
Baber's avatar
Baber committed
72
73

    @cached_property
Baber's avatar
Baber committed
74
    def metric_name(self) -> str:
Baber's avatar
Baber committed
75
76
77
78
79
80
81
82
83
84
85
86
87
88
        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
89
    def compute_metric(self, *args, **kwargs) -> Any:
Baber's avatar
Baber committed
90
91
92
93
94
        """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
95
96
97
98
99
100
    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
101

102
103
104
105
106
@dataclass
class RepeatConfig:
    """Encapsulates information about a single repeat."""

    repeats: int = 1
Baber's avatar
Baber committed
107
    metric_fn: Optional[Callable] = "pass@N"
108
109
110
    kwargs: Optional[dict] = None


Baber's avatar
Baber committed
111
112
@dataclass
class FilterConfig:
Baber's avatar
nit  
Baber committed
113
    """Encapsulates information about a single filter."""
Baber's avatar
Baber committed
114
115
116
117
118
119

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


Baber's avatar
Baber committed
120
121
122
123
@dataclass
class FewshotConfig:
    sampler: str
    samples: list[dict]
124
    process_docs: Optional[Callable] = None
Baber's avatar
Baber committed
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
    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:
Baber's avatar
Baber committed
152
153
154
155
156
157
158
159
160
    """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.
    """
Baber's avatar
Baber committed
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178

    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:
Baber's avatar
Baber committed
179
180
181
182
183
    """Encapsulates information about a template.
    Would return a sample with the following format:
    Question:  <doc_to_text(doc)>
    Answer:` <doc_to_target(doc)>`
    """
Baber's avatar
Baber committed
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200

    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
201
202


Baber's avatar
Baber committed
203
204
205
206
207
208
209
210
211
212
@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


213
214
@dataclass
class TaskConfig(dict):
215
    # task naming/registry
216
217
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
218
    tag: Optional[Union[str, list]] = None
219
220
221
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
222
    custom_dataset: Optional[Callable] = None
223
224
225
226
227
228
    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
229
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
230
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
231
    )
232
233
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
234
235
236
    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
237
238
    doc_to_image: Union[Callable, str, None] = None
    doc_to_audio: Union[Callable, str, None] = None
Hojin Lee's avatar
Hojin Lee committed
239
    unsafe_code: bool = False
240
241
242
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
243
    description: str = ""
244
245
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
246
    fewshot_config: Optional[dict] = None
247
    # runtime configuration options
248
    num_fewshot: Optional[int] = None
249
    # scoring options
250
251
252
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
253
    repeats: int = 1
Baber's avatar
Baber committed
254
    filter_list: Optional[list[dict]] = None
255
    should_decontaminate: bool = False
256
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
257
    gen_prefix: Optional[str] = None
258
259
260
    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
261
262
    _metric_list: list[MetricConfig] = None
    _filter_list: list[FilterConfig] = None
263

Ethan Smith's avatar
Ethan Smith committed
264
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
265
        if self.generation_kwargs is not None:
266
            if self.output_type != "generate_until":
267
                eval_logger.warning(
268
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
269
270
271
272
273
274
275
276
                )

            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
277
278
279
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
280
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
281
        else:
282
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
283
284
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
285
286
287
288
289
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
290
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
291
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
292
                }
Baber Abbasi's avatar
Baber Abbasi committed
293
294
295
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
296

Baber's avatar
Baber committed
297
298
        if self.metric_list and not all("metric" in cfg for cfg in self.metric_list):
            raise ValueError("each entry in metric_list must include a 'metric' key")
Baber's avatar
Baber committed
299
300
301
302

    def get_metrics(self) -> list["MetricConfig"]:
        metrics = []
        if self.metric_list is None:
Baber's avatar
Baber committed
303
            # ---------- 1. If no metrics defined, use defaults for output type ----------
Baber's avatar
Baber committed
304
            _metric_list = DEFAULT_METRIC_REGISTRY[self.output_type]
Baber's avatar
Baber committed
305
306
307
            eval_logger.info(
                f"No metrics defined in config, using default metrics for {self.output_type}={_metric_list}"
            )
Baber's avatar
Baber committed
308
309
310
311
312
313
314
315
316
317
            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:
Baber's avatar
Baber committed
318
            # ---------- 2. How will the samples be evaluated ----------
Baber's avatar
Baber committed
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
            for metric_config in self.metric_list:
                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
381
382
    def get_filters(self) -> list["FilterEnsemble"]:
        if not self.filter_list:
Baber's avatar
Baber committed
383
            eval_logger.debug(
Baber's avatar
Baber committed
384
                "No custom filters defined; falling back to 'take_first' for handling repeats."
Baber's avatar
Baber committed
385
            )
Baber's avatar
Baber committed
386
387
            return [build_filter_ensemble("none", [["take_first", None]])]
        else:
Baber's avatar
Baber committed
388

Baber's avatar
Baber committed
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
            def _strip_fn(d: dict) -> dict:
                return {k: v for k, v in d.items() if k != "function"}

            configs = (
                self.filter_list.values()
                if isinstance(self.filter_list, dict)
                else self.filter_list
            )

            return [
                build_filter_ensemble(
                    filter_name=cfg["name"],
                    components=[[_strip_fn(f) for f in cfg["filter"]]],
                )
                for cfg in configs
            ]
Baber's avatar
Baber committed
405

406
407
408
    def __getitem__(self, item):
        return getattr(self, item)

409
410
411
    def __setitem__(self, item, value):
        return setattr(self, item, value)

412
    def to_dict(self, keep_callable: bool = False) -> dict:
Baber's avatar
Baber committed
413
        """Return a printable dict with Nones stripped and callables serialised.
414

haileyschoelkopf's avatar
haileyschoelkopf committed
415
416
417
        :return: dict
            A printable dictionary version of the TaskConfig object.
        """
Baber's avatar
Baber committed
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433

        def _maybe_serialize(val):
            return (
                self.serialize_function(val, keep_callable=keep_callable)
                if callable(val)
                else val
            )

        cfg = asdict(self)
        return {
            k: [{mk: _maybe_serialize(mv) for mk, mv in md.items()} for md in v]
            if k == "metric_list"
            else _maybe_serialize(v)
            for k, v in cfg.items()
            if v is not None
        }
434

435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
    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)

451
452
453
454
455
456
457
458
459
460
461

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)
    """

462
    VERSION: Optional[Union[int, str]] = None
463

464
465
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
466
    DATASET_PATH: Optional[str] = None
467
468

    # The name of a subset within `DATASET_PATH`.
469
    DATASET_NAME: Optional[str] = None
470

471
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
472

473
474
    def __init__(
        self,
475
476
477
478
        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
479
    ) -> None:
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
        """
        :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)
502
503
504
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
505

506
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
507

lintangsutawika's avatar
lintangsutawika committed
508
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
509
510
511
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
512

513
514
515
516
517
518
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
        """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.
        """
543
544
545
546
547
548
549
        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,
        )
550

551
    @property
552
    def config(self) -> TaskConfig:
553
554
555
        """Returns the TaskConfig associated with this class."""
        return self._config

556
    @abc.abstractmethod
Baber's avatar
Baber committed
557
    def has_training_docs(self) -> bool:
558
559
560
561
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
562
    def has_validation_docs(self) -> bool:
563
564
565
566
        """Whether the task has a validation set"""
        pass

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

571
    def training_docs(self) -> Iterable:
572
573
574
575
576
577
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

578
    def validation_docs(self) -> Iterable:
579
580
581
582
583
584
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

585
    def test_docs(self) -> Iterable:
586
587
588
589
590
591
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

592
    def fewshot_docs(self) -> Iterable:
593
594
595
596
597
598
599
600
601
        """
        :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
602
603
604
605
606
            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."
                )
607
608
            return self.test_docs()

609
    def _process_doc(self, doc: dict) -> dict:
610
611
612
613
614
615
616
617
618
        """
        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
619

620
    @property
Baber's avatar
Baber committed
621
    def instances(self) -> list[Instance]:
622
623
624
625
626
        """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
627
    def fewshot_examples(self, k, rnd) -> Iterable[dict]:
628
629
630
631
632
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

Baber's avatar
Baber committed
633
    def doc_to_decontamination_query(self, doc: dict):
634
        raise NotImplementedError(
635
636
637
638
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
Baber committed
639
    def doc_to_text(self, doc: dict) -> str:
640
641
642
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
643
    def doc_to_target(self, doc: dict) -> Union[str, int]:
644
645
        pass

646
    # not an abstractmethod because not every language-only task has to implement this
Baber's avatar
Baber committed
647
    def doc_to_image(self, doc: dict):
648
649
        raise NotImplementedError

Baber's avatar
Baber committed
650
    def doc_to_audio(self, doc: dict):
651
652
        raise NotImplementedError

Baber's avatar
Baber committed
653
    def doc_to_prefix(self, doc: dict) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
654
655
        return ""

656
657
    def build_all_requests(
        self,
658
        *,
659
        limit: Union[int, None] = None,
660
        samples: Optional[List[int]] = None,
661
662
663
664
665
666
667
668
669
        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 = "",
670
    ) -> None:
671
        """Build a set of Instances for a task, and store them in task.instances"""
672
673
674
675

        # used with caching
        og_limit = limit

676
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
677
678
679
680
681
682
683
        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 ""
        )
684
        cache_key += f"-tokenizer{tokenizer_name}"
685

Baber Abbasi's avatar
Baber Abbasi committed
686
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
687
688
689
690
691
692
693
694
695
696
697
698
699

        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
700
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
701

702
        instances = []
703
704
705
706
707
708
709
710
711
712

        # 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(
713
714
715
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
716
717
718
719
720
721
722
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
723
        ):
724
            # sample fewshot context #TODO: need to offset doc_id by rank now!
725
            fewshot_ctx = self.fewshot_context(
726
                doc,
727
728
729
730
731
732
733
                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
734
                gen_prefix=self.doc_to_prefix(doc),
735
            )
736

737
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
738
739
740
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
741
                metadata=(self.config["task"], doc_id, self.config.repeats),
742
                apply_chat_template=apply_chat_template,
743
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
744
            )
745
746
747
748

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

749
750
751
752
753
754
755
756
757
758
759
760
761
            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
762

763
764
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
765

766
767
768
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

769
    @abc.abstractmethod
Baber's avatar
Baber committed
770
    def construct_requests(self, doc: dict, ctx: Union[list[dict], str], **kwargs):
771
772
773
774
775
776
777
778
779
780
781
782
783
784
        """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
785
            The number of times each instance in a dataset is inferred on. Defaults to 1,
786
787
788
789
790
            can be increased for techniques like majority voting.
        """
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
791
    def process_results(self, doc: dict, results: list):
792
793
794
795
796
797
798
799
800
801
802
        """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
803
    @deprecated("not used anymore")
804
805
806
807
808
809
810
811
    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
812
    @deprecated("not used anymore")
813
814
815
816
817
818
819
820
    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

821
822
823
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
824
    @classmethod
Baber's avatar
Baber committed
825
    def count_bytes(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
826
827
828
829
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
830
    def count_words(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
831
832
833
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

834
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
835
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
836
837
838
839
840
841
842
        """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
843
844
845
846
847
        :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.
848
849
850
        :returns: str
            The fewshot context.
        """
851
        if rnd is None:
852
853
854
855
856
857
            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
858

859
        description = description if description else ""
860
861

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
862
            labeled_examples = ""
863
        else:
lintangsutawika's avatar
lintangsutawika committed
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
            # 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
888
            )
889
890

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
891
        return description + labeled_examples + example
892

893
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
894
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
895
896
        if hasattr(self, "_filters"):
            for f in self._filters:
897
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
898
899
900
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
901

baberabb's avatar
baberabb committed
902
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
903
        """Returns the config as a dictionary."""
904
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
905
        # (num_fewshot)
906
        return self.config.to_dict()
907

Baber Abbasi's avatar
Baber Abbasi committed
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
    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
930
931
932
933
934
935
936
        # 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
937

938
939
940
941
942
    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

943
    @property
Baber's avatar
Baber committed
944
    def eval_docs(self) -> Union[datasets.Dataset, Iterable[dict]]:
945
946
947
948
949
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
950
951
952
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
953
954

    def doc_iterator(
955
956
957
958
959
960
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
961
    ) -> Iterator[Tuple[int, Any]]:
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
        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),
            )
984
985
        return doc_iterator

986
987

class ConfigurableTask(Task):
988
    VERSION = "Yaml"
989
    OUTPUT_TYPE = None
990
    CONFIG = None
991
992

    def __init__(
993
994
995
996
997
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Baber's avatar
Baber committed
998
    ) -> None:
999
        # Get pre-configured attributes
1000
        self._config = self.CONFIG
1001

1002
        # Use new configurations if there was no preconfiguration
1003
        if self.config is None:
1004
            self._config = TaskConfig(**config)
1005
1006
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
1007
            if config is not None:
1008
                self._config.__dict__.update(config)
1009

1010
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
1011
1012
1013
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
1014

1015
1016
1017
1018
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

1019
        if self.config.output_type is not None:
1020
1021
1022
1023
            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)}'"
                )
1024
            self.OUTPUT_TYPE = self.config.output_type
1025

1026
1027
1028
1029
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

1030
1031
1032
1033
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
1034
1035
1036
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

1037
1038
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
1039

1040
1041
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
1042

Baber's avatar
Baber committed
1043
        self.metric_list: list[MetricConfig] = self.config.get_metrics()
1044

1045
        self.download(self.config.dataset_kwargs)
1046
1047
1048
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
1049
1050
        self._filters = self.config.get_filters()

1051
1052
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
1053
            self.prompt = get_prompt(
1054
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
1055
            )
1056
1057
1058
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
1059
        if self.fewshot_docs() is not None:
1060
1061
1062
1063
            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
1064
1065
1066
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
            )
            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)}"
                )
1083

1084
        self.task_docs = self.eval_docs
1085

1086
        # Test One Doc
1087
        self.features = list(self.task_docs.features.keys())
1088
1089
        self.multiple_input = 0
        self.multiple_target = 0
1090
        test_doc = self.task_docs[0]
1091
        test_text = self.doc_to_text(test_doc)
1092
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
1093

1094
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1095
            test_choice = self.doc_to_choice(test_doc)
1096
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
1097
                eval_logger.error("doc_to_choice must return list")
1098
1099
            else:
                num_choice = len(test_choice)
1100

1101
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
1102
1103
1104
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
1105
                self.multiple_input = num_choice
1106
1107
        else:
            test_choice = None
1108

1109
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
1110
1111
1112
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
1113
            self.multiple_target = len(test_target)
1114
        else:
1115
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1116
                test_target = test_choice[test_target]
1117
            else:
lintangsutawika's avatar
lintangsutawika committed
1118
                test_target = str(test_target)
1119

1120
1121
1122
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1123
            check_choices = [test_target]
1124
1125
1126
1127
        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 = (
1128
1129
                    True
                    if self.config.target_delimiter.rstrip()
1130
                    != self.config.target_delimiter
1131
                    else False
1132
                )
1133

1134
                if delimiter_has_whitespace and choice_has_whitespace:
1135
1136
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1137
1138
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1139
                    eval_logger.debug(
1140
                        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'
1141
1142
                    )

Baber Abbasi's avatar
Baber Abbasi committed
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
        if isinstance(self.config.custom_dataset, Callable):
            eval_logger.warning(
                f"{self.config.task}: Custom kwargs can be passed to `--metadata` in console (as json string) or to the TaskManager."
                + "\nFor example --metadata='{\"max_seq_lengths\":[4096, 8192]}'. For details see task Readme."
            )
            self.dataset = self.config.custom_dataset(
                **(self.config.metadata or {}), **(self.config.dataset_kwargs or {})
            )
        else:
            self.dataset = datasets.load_dataset(
                path=self.DATASET_PATH,
                name=self.DATASET_NAME,
                **dataset_kwargs if dataset_kwargs is not None else {},
            )
1160

baberabb's avatar
baberabb committed
1161
    def has_training_docs(self) -> bool:
1162
        if self.config.training_split is not None:
1163
1164
1165
1166
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1167
    def has_validation_docs(self) -> bool:
1168
        if self.config.validation_split is not None:
1169
1170
1171
1172
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1173
    def has_test_docs(self) -> bool:
1174
        if self.config.test_split is not None:
1175
1176
1177
1178
            return True
        else:
            return False

Baber's avatar
Baber committed
1179
    def training_docs(self) -> Optional[datasets.Dataset]:
1180
        if self.has_training_docs():
1181
1182
1183
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1184
                )
1185
            return self.dataset[self.config.training_split]
1186

Baber's avatar
Baber committed
1187
    def validation_docs(self) -> Optional[datasets.Dataset]:
1188
        if self.has_validation_docs():
1189
1190
1191
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1192
                )
1193
            return self.dataset[self.config.validation_split]
1194

Baber's avatar
Baber committed
1195
    def test_docs(self) -> Optional[datasets.Dataset]:
1196
        if self.has_test_docs():
1197
1198
1199
            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]
1200

1201
    def fewshot_docs(self):
1202
        if self.config.fewshot_split is not None:
1203
1204
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1205
            return self.dataset[self.config.fewshot_split]
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
        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."
                )
1218
        else:
1219
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1220
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1221
                    f"[Task: {self.config.task}] "
1222
1223
1224
1225
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1226

KonradSzafer's avatar
KonradSzafer committed
1227
1228
1229
1230
1231
    @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
1232
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1233
1234
1235
1236
1237
1238
1239
1240
    ) -> 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":
1241
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1242
1243
            # if last message is user, append to it to avoid two user messages in a row
            else:
1244
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1245
1246
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1247
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1248
1249
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1250

lintangsutawika's avatar
lintangsutawika committed
1251
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1252
1253
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1254
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1255
1256
1257
1258
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1259
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1260
        gen_prefix: Optional[str] = None,
Baber's avatar
Baber committed
1261
    ) -> Union[str, List[str], None]:
lintangsutawika's avatar
lintangsutawika committed
1262
1263
1264
1265
1266
1267
1268
        """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
1269
1270
1271
1272
1273
1274
        :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.
1275
1276
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1277
1278
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1279
1280
1281
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1282
1283
1284
1285
1286
1287
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1288
1289
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1290

KonradSzafer's avatar
KonradSzafer committed
1291
1292
1293
1294
1295
1296
1297
1298
1299
        # 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
1300
        else:
KonradSzafer's avatar
KonradSzafer committed
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
            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
1314
1315
1316
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1317
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1318
1319
1320
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1321
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1322
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1323
                )
lintangsutawika's avatar
lintangsutawika committed
1324
1325

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1326
1327
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1328
                # TODO: append prefill?
1329
1330
                if not labeled_examples:
                    return ""
1331
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1332
1333
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1334
1335
1336
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1337
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1338
1339
1340
1341
1342
1343
1344
                )
            # 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
1345
1346
1347
1348
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1349
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1350
1351
1352
1353
1354
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1355
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1356
1357
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1358
1359
1360
1361
1362
1363
                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
1364
1365
1366
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1367
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1368
1369
1370
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1371
1372
1373
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1374
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1375
1376
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1377
1378
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1379
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1380
            )
1381
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1382
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1383
1384
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1385
1386
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1387
1388
            if self.multiple_input:
                return labeled_examples
1389
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1390
                return labeled_examples + example + prefix
1391
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1392
                return [labeled_examples + ex + prefix for ex in example]
1393
1394
1395
            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
1396
                    return labeled_examples + choices[example] + prefix
1397
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1398
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1399

Baber Abbasi's avatar
Baber Abbasi committed
1400
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1401
        """Iterates over FilterEnsembles and applies them to instances"""
1402
1403
        if hasattr(self, "_filters"):
            for f in self._filters:
1404
                f.apply(self._instances)
1405
1406
1407
1408
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1409
    def should_decontaminate(self):
1410
        return self.config.should_decontaminate
1411

Baber Abbasi's avatar
Baber Abbasi committed
1412
    def doc_to_decontamination_query(self, doc: dict):
1413
        if self.config.should_decontaminate:
1414
1415
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1416
            else:
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
                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
                        )
                    )
1428

1429
    def _process_doc(self, doc: dict) -> dict:
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

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

Baber's avatar
Baber committed
1440
    def doc_to_text(self, doc: dict, doc_to_text: Optional[int, str, Callable] = None):
1441
1442
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1443
1444
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1445
        else:
1446
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1447

1448
        if isinstance(doc_to_text, int):
1449
            return doc_to_text
1450
        elif isinstance(doc_to_text, str):
1451
            if doc_to_text in self.features:
1452
                # if self.config.doc_to_choice is not None:
1453
1454
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1455
1456
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1457
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
nit  
Baber committed
1458
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1459
1460
1461
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1462
        elif callable(doc_to_text):
1463
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1464
        # Used when applying a Promptsource template
1465
        elif hasattr(doc_to_text, "apply"):
1466
1467
1468
1469
1470
            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")
1471
                return self.config.fewshot_delimiter
1472
        else:
1473
            print(type(doc_to_text))
1474
            raise TypeError
1475

Baber's avatar
Baber committed
1476
    def doc_to_target(self, doc: dict, doc_to_target=None) -> Union[int, str, list]:
1477
1478
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1479
1480
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1481
        else:
1482
            doc_to_target = self.config.doc_to_target
1483

1484
        if isinstance(doc_to_target, int):
1485
            return doc_to_target
1486
        elif isinstance(doc_to_target, str):
1487
            if doc_to_target in self.features:
1488
                # if self.config.doc_to_choice is not None:
1489
1490
1491
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1492
            else:
lintangsutawika's avatar
lintangsutawika committed
1493
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
nit  
Baber committed
1494
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1495
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1496
1497
1498
1499
1500
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1501
1502
1503
1504
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1505
1506
                else:
                    return target_string
1507
        elif isinstance(doc_to_target, list):
1508
            return doc_to_target
1509
        elif callable(doc_to_target):
1510
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1511
        # Used when applying a Promptsource template
1512
        elif hasattr(doc_to_target, "apply"):
1513
            applied_prompt = doc_to_target.apply(doc)
1514
1515
1516
1517
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1518
                return self.config.fewshot_delimiter
1519
1520
        else:
            raise TypeError
1521

Baber's avatar
Baber committed
1522
1523
1524
    def doc_to_choice(
        self, doc: dict, doc_to_choice: Union[str, list, dict] = None
    ) -> List[str]:
1525
1526
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1527
1528
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1529
        elif self.config.doc_to_choice is None:
1530
1531
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1532
            doc_to_choice = self.config.doc_to_choice
1533

1534
        if isinstance(doc_to_choice, str):
1535
1536
1537
1538
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1539
        elif isinstance(doc_to_choice, list):
1540
            return doc_to_choice
1541
        elif isinstance(doc_to_choice, dict):
1542
1543
1544
1545
1546
1547
1548
            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
1549

Baber's avatar
Baber committed
1550
    def doc_to_image(self, doc: dict, doc_to_image=None) -> Union[int, str, list, None]:
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
        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
1573
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list, None]:
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
        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
1596
    def doc_to_prefix(self, doc: dict) -> Optional[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1597
1598
1599
1600
1601
1602
1603
        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
1604
1605
1606
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1607
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1608
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1609

1610
1611
        aux_arguments = None

1612
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1613
            arguments = (ctx, self.doc_to_target(doc))
1614
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1615
            arguments = (self.doc_to_target(doc),)
1616
        elif self.OUTPUT_TYPE == "multiple_choice":
1617
            choices = self.doc_to_choice(doc)
1618
            target_delimiter = self.config.target_delimiter
1619
1620
            if apply_chat_template:
                target_delimiter = ""
1621
1622
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1623
                # apply chat_template to choices if apply_chat_template
1624
                cont = self.doc_to_target(doc)
1625

1626
                arguments = [
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1637
                ]
1638
            else:
1639
                # Otherwise they are placed in the continuation
1640
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1641

1642
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
Baber's avatar
Baber committed
1643
            if "acc_mutual_info" in [m.metric_name for m in self.metric_list]:
1644
1645
1646
1647
1648
1649
                # 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.
1650
1651
1652
1653
                # 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
                ]
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668

                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)},
            }

1669
1670
1671
1672
1673
1674
1675
1676
        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)},
            }

1677
1678
1679
1680
1681
1682
1683
        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":
1684
            request_list = [
1685
1686
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1687
                    doc=doc,
1688
                    arguments=arg,
1689
                    idx=i,
1690
1691
                    **kwargs,
                )
1692
                for i, arg in enumerate(arguments)
1693
            ]
1694
1695

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1696

lintangsutawika's avatar
lintangsutawika committed
1697
        return Instance(
1698
1699
1700
1701
1702
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1703
        )
1704

Baber's avatar
Baber committed
1705
    def process_results(self, doc: dict, results: list) -> dict:
1706
1707
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1708

1709
        result_dict = {}
Baber's avatar
Baber committed
1710
        use_metric = list(m.metric_name for m in self.metric_list)
1711
1712
1713
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1714
1715
1716
1717
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1718
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1719
            (loglikelihood,) = results
1720
1721
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1722
            return {
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
1736
1737
                **(
                    {"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
1738
            }
1739
        elif self.OUTPUT_TYPE == "multiple_choice":
1740
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1741

1742
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1743
            choices = self.doc_to_choice(doc)
1744
1745
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1746
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1747
1748
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1749
1750
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1751
1752
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1753
                # and this stores our "regular" conditional loglikelihoods
1754
                lls = lls[: len(choices)]
1755

1756
1757
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1758

1759
1760
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1761
            else:
1762
                gold = self.doc_to_target(doc)
1763
1764

            gold_index_error = False
1765
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1766
1767
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1768
1769
                    gold_index_error = True
            else:
1770
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1771
                    gold = gold if gold < len(choices) else -100
1772
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1773
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1774

Lintang Sutawika's avatar
Lintang Sutawika committed
1775
                if gold == -100:
1776
1777
1778
1779
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1780
                    f"Label index was not in within range of available choices,"
1781
1782
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1783

1784
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1785
1786
                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
1787
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1788
1789
1790
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1791
                # 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
1792
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1793

Lintang Sutawika's avatar
Lintang Sutawika committed
1794
1795
1796
1797
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1798
            result_dict = {
1799
                **({"acc": acc} if "acc" in use_metric else {}),
1800
1801
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1802
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1803
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1804
1805
1806
1807
1808
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1809
1810
            }

1811
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1812
1813
1814
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1815
1816
1817
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1818
        elif self.OUTPUT_TYPE == "generate_until":
1819
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1820
            result = results[0]
1821
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1822
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1823
                # it assumes that doc_to_target returns a number.
1824
1825
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1826
1827
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1828
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1829
1830
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
Baber's avatar
Baber committed
1831
                "bypass" in use_metric or isinstance(result, list)
1832
            ):
Chris's avatar
Chris committed
1833
1834
                # cast gold to the same type as result
                gold = type(result)(gold)
1835

Baber's avatar
Baber committed
1836
            for metric in self.metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
1837
1838
1839
1840
1841
                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
1842
1843
1844
1845
                    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
1846
                    if metric.name == "exact_match":
1847
                        result = [result for _ in range(len(gold))]
Baber's avatar
Baber committed
1848
                        scores = metric.fn(
1849
1850
                            references=gold,
                            predictions=result,
Baber's avatar
Baber committed
1851
                            **metric.kwargs,
1852
1853
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1854
                    else:
1855
1856
                        for gold_option in gold:
                            try:
Baber's avatar
Baber committed
1857
                                result_score = metric.fn(
1858
1859
                                    references=[gold_option],
                                    predictions=[result],
Baber's avatar
Baber committed
1860
                                    **metric.kwargs,
1861
1862
1863
1864
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
Baber's avatar
Baber committed
1865
                                result_score = metric.fn([gold_option, result])
1866
1867
1868
1869
1870
1871
1872
1873
                            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
1874
                else:
1875
                    try:
Baber's avatar
Baber committed
1876
                        result_score = metric.fn(
1877
1878
                            references=[gold],
                            predictions=[result],
Baber's avatar
Baber committed
1879
                            **metric.kwargs,
1880
                        )
1881
                    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
1882
                        result_score = metric.fn([gold, result])
1883
1884
1885
1886
1887
1888
1889
                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
1890
        else:
lintangsutawika's avatar
lintangsutawika committed
1891
1892
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1893
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1894
            )
1895
1896
1897

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1898
    def aggregation(self) -> dict:
Baber's avatar
Baber committed
1899
        return {k.name: k.aggregation_fn for k in self.metric_list}
1900

Baber Abbasi's avatar
Baber Abbasi committed
1901
    def higher_is_better(self) -> dict:
Baber's avatar
Baber committed
1902
        return {k.name: k.higher_is_better for k in self.metric_list}
1903

Baber Abbasi's avatar
Baber Abbasi committed
1904
1905
1906
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1907
    @property
Baber's avatar
Baber committed
1908
    def task_name(self) -> Optional[str]:
Lintang Sutawika's avatar
Lintang Sutawika committed
1909
1910
        return getattr(self.config, "task", None)

1911
1912
1913
1914
1915
    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
1916
            f"num_samples={len(self.eval_docs)})"
1917
1918
        )

1919
1920

class MultipleChoiceTask(Task):
1921
    OUTPUT_TYPE = "loglikelihood"
1922

baberabb's avatar
baberabb committed
1923
    def doc_to_target(self, doc: dict) -> str:
1924
1925
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1926
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1927
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1928
1929
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1930
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1931
                doc=doc,
1932
                arguments=(ctx, " {}".format(choice)),
1933
                idx=i,
1934
1935
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1936
1937
            for i, choice in enumerate(doc["choices"])
        ]
1938

1939
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1940
1941
1942
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
        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
1954
    def higher_is_better(self) -> dict:
1955
1956
1957
1958
1959
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1960
    def aggregation(self) -> dict:
1961
1962
1963
1964
1965
1966
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1967
class PerplexityTask(Task):
1968
1969
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1970
    def has_training_docs(self) -> bool:
1971
1972
        return False

baberabb's avatar
baberabb committed
1973
    def fewshot_examples(self, k: int, rnd) -> List:
1974
1975
1976
1977
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1978
1979
        return []

baberabb's avatar
baberabb committed
1980
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1981
1982
1983
1984
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1985
1986
1987

        return ""

baberabb's avatar
baberabb committed
1988
    def higher_is_better(self) -> dict:
1989
1990
1991
1992
1993
1994
1995
1996
1997
        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
1998
    def doc_to_text(self, doc) -> str:
1999
2000
2001
2002
2003
        return ""

    def doc_to_target(self, doc):
        return doc

2004
2005
2006
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
2007

lintangsutawika's avatar
lintangsutawika committed
2008
2009
2010
2011
2012
2013
2014
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
2015

2016
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
2017
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
2018
2019
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
2020
2021
2022
2023
2024
2025
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
2026
    def aggregation(self) -> dict:
2027
2028
2029
2030
2031
2032
2033
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
2034
    def count_bytes(cls, doc) -> int:
2035
2036
2037
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
2038
    def count_words(cls, doc) -> int:
2039
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
2040
        return len(re.split(r"\s+", doc))