task.py 74.1 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
8
from dataclasses import asdict, dataclass
Baber's avatar
Baber committed
9
from functools import cached_property
10
from inspect import getsource
11
12
13
14
15
16
17
18
19
20
21
22
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
23
24
25

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

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

45

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

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

lintangsutawika's avatar
lintangsutawika committed
55

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

    name: str
    fn: Optional[Callable] = None
    kwargs: Optional[dict] = None
    aggregation_fn: Optional[Callable] = None
    higher_is_better: bool = True
    hf_evaluate: bool = False
66
67
    sample_metric: bool = True
    is_elementwise: bool = True
Baber's avatar
Baber committed
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85

    @cached_property
    def metric_names(self) -> str:
        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


86
87
88
89
90
91
92
93
94
@dataclass
class RepeatConfig:
    """Encapsulates information about a single repeat."""

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


Baber's avatar
Baber committed
95
96
97
98
99
100
101
102
103
@dataclass
class FilterConfig:
    """Encapsulates information about a single filter."""

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


Baber's avatar
Baber committed
104
105
106
107
@dataclass
class FewshotConfig:
    sampler: str
    samples: list[dict]
108
    process_docs: Optional[Callable] = None
Baber's avatar
Baber committed
109
110


111
112
@dataclass
class TaskConfig(dict):
113
    # task naming/registry
114
115
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
116
    tag: Optional[Union[str, list]] = None
117
118
119
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
120
    custom_dataset: Optional[Callable] = None
121
122
123
124
125
126
    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
127
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
128
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
129
    )
130
131
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
132
133
134
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
135
    doc_to_image: Union[Callable, str] = None
136
    doc_to_audio: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
137
    unsafe_code: bool = False
138
139
140
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
141
    description: str = ""
142
143
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
144
    fewshot_config: Optional[dict] = None
145
    # runtime configuration options
146
    num_fewshot: Optional[int] = None
147
    # scoring options
148
149
150
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
151
    repeats: int = 1
152
    filter_list: Optional[Union[str, list]] = None
153
    should_decontaminate: bool = False
154
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
155
    gen_prefix: Optional[str] = None
156
157
158
    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
159
160
    _metric_list = None
    _filter_list = None
161

Ethan Smith's avatar
Ethan Smith committed
162
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
163
        if self.generation_kwargs is not None:
164
            if self.output_type != "generate_until":
165
                eval_logger.warning(
166
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
167
168
169
170
171
172
173
174
                )

            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
175
176
177
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
178
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
179
        else:
180
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
181
182
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
183
184
185
186
187
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
188
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
189
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
190
                }
Baber Abbasi's avatar
Baber Abbasi committed
191
192
193
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
194

Baber's avatar
Baber committed
195
196
197
198
199
200
201
202
203
204
205
        if self.metric_list is not None:
            for metric_config in self.metric_list:
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )

    def get_metrics(self) -> list["MetricConfig"]:
        metrics = []
        if self.metric_list is None:
            _metric_list = DEFAULT_METRIC_REGISTRY[self.output_type]
Baber's avatar
Baber committed
206
207
208
            eval_logger.info(
                f"No metrics defined in config, using default metrics for {self.output_type}={_metric_list}"
            )
Baber's avatar
Baber committed
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
            metrics.extend(
                MetricConfig(
                    name=metric_name,
                    fn=get_metric(metric_name),
                    aggregation_fn=get_metric_aggregation(metric_name),
                    higher_is_better=is_higher_better(metric_name),
                )
                for metric_name in _metric_list
            )
        else:
            for metric_config in self.metric_list:
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
                metric_name = metric_config["metric"]
                _metric_fn_kwargs = {
                    key: metric_config[key]
                    for key in metric_config
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
                }
                _hf_evaluate_metric: bool = metric_config.get("hf_evaluate", False)
                _metric_fn = None
                _aggregation = None

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

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

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

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

Baber's avatar
Baber committed
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
    def get_filters(self):
        if self.filter_list is not None:
            _filter_list = []
            if isinstance(self.filter_list, dict):
                for filter_config in self.filter_list:
                    _filter_list.append(
                        build_filter_ensemble(
                            filter_name=filter_config["name"],
                            components=[
                                [
                                    {
                                        key: function[key]
                                        for key in function
                                        if key != "function"
                                    }
                                ]
                                for function in filter_config["filter"]
                            ],
                        )
                    )
        else:
            # TODO: handle repeats in a more general way rather than just discarding
            eval_logger.debug(
                "No custom filters defined. Using default 'take_first' filter for handling repeats."
            )
            _filter_list = [build_filter_ensemble("none", [["take_first", None]])]

        return _filter_list

314
315
316
    def __getitem__(self, item):
        return getattr(self, item)

317
318
319
    def __setitem__(self, item, value):
        return setattr(self, item, value)

320
    def to_dict(self, keep_callable: bool = False) -> dict:
321
322
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
323
        Used for dumping results alongside full task configuration
324

haileyschoelkopf's avatar
haileyschoelkopf committed
325
326
327
328
329
330
331
332
333
334
        :return: dict
            A printable dictionary version of the TaskConfig object.

        # TODO: should any default value in the TaskConfig not be printed?
        """
        cfg_dict = asdict(self)
        # remove values that are `None`
        for k, v in list(cfg_dict.items()):
            if v is None:
                cfg_dict.pop(k)
335
336
337
338
339
340
341
342
343
344
            elif k == "metric_list":
                for metric_dict in v:
                    for metric_key, metric_value in metric_dict.items():
                        if callable(metric_value):
                            metric_dict[metric_key] = self.serialize_function(
                                metric_value, keep_callable=keep_callable
                            )
                cfg_dict[k] = v
            elif callable(v):
                cfg_dict[k] = self.serialize_function(v, keep_callable=keep_callable)
haileyschoelkopf's avatar
haileyschoelkopf committed
345
        return cfg_dict
346

347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
    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)

363
364
365
366
367
368
369
370
371
372
373

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)
    """

374
    VERSION: Optional[Union[int, str]] = None
375

376
377
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
378
    DATASET_PATH: Optional[str] = None
379
380

    # The name of a subset within `DATASET_PATH`.
381
    DATASET_NAME: Optional[str] = None
382

383
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
384

385
386
    def __init__(
        self,
387
388
389
390
        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
391
    ) -> None:
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
        """
        :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)
414
415
416
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
417

418
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
419

lintangsutawika's avatar
lintangsutawika committed
420
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
421
422
423
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
424

425
426
427
428
429
430
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
        """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.
        """
455
456
457
458
459
460
461
        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,
        )
462

463
    @property
464
    def config(self) -> TaskConfig:
465
466
467
        """Returns the TaskConfig associated with this class."""
        return self._config

468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

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

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

483
    def training_docs(self) -> Iterable:
484
485
486
487
488
489
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

490
    def validation_docs(self) -> Iterable:
491
492
493
494
495
496
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

497
    def test_docs(self) -> Iterable:
498
499
500
501
502
503
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

504
    def fewshot_docs(self) -> Iterable:
505
506
507
508
509
510
511
512
513
        """
        :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
514
515
516
517
518
            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."
                )
519
520
            return self.test_docs()

521
    def _process_doc(self, doc: dict) -> dict:
522
523
524
525
526
527
528
529
530
        """
        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
531

532
    @property
533
    def instances(self) -> List[Instance]:
534
535
536
537
538
539
540
541
542
543
544
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

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

        return rnd.sample(self._training_docs, k)

545
546
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
547
548
549
550
551
552
553
554
555
556
557
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

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

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

558
559
560
561
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

562
563
564
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
565
566
567
    def doc_to_prefix(self, doc):
        return ""

568
569
    def build_all_requests(
        self,
570
        *,
571
        limit: Union[int, None] = None,
572
        samples: Optional[List[int]] = None,
573
574
575
576
577
578
579
580
581
        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 = "",
582
    ) -> None:
583
        """Build a set of Instances for a task, and store them in task.instances"""
584
585
586
587

        # used with caching
        og_limit = limit

588
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
589
590
591
592
593
594
595
        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 ""
        )
596
        cache_key += f"-tokenizer{tokenizer_name}"
597

Baber Abbasi's avatar
Baber Abbasi committed
598
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
599
600
601
602
603
604
605
606
607
608
609
610
611

        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
612
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
613

614
        instances = []
615
616
617
618
619
620
621
622
623
624

        # 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(
625
626
627
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
628
629
630
631
632
633
634
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
635
        ):
636
            # sample fewshot context #TODO: need to offset doc_id by rank now!
637
            fewshot_ctx = self.fewshot_context(
638
                doc,
639
640
641
642
643
644
645
                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
646
                gen_prefix=self.doc_to_prefix(doc),
647
            )
648

649
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
650
651
652
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
653
                metadata=(self.config["task"], doc_id, self.config.repeats),
654
                apply_chat_template=apply_chat_template,
655
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
656
            )
657
658
659
660

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

661
662
663
664
665
666
667
668
669
670
671
672
673
            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
674

675
676
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
677

678
679
680
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
    @abc.abstractmethod
    def construct_requests(self, doc, ctx, **kwargs):
        """Uses RequestFactory to construct Requests and returns an iterable of
        Requests which will be sent to the LM.

        :param doc:
            The document as returned from training_docs, validation_docs, or test_docs.
        :param ctx: str
            The context string, generated by fewshot_context. This includes the natural
            language description, as well as the few shot examples, and the question
            part of the document for `doc`.
        :param doc_idx: int
            The index of a document within `self.test_docs()` or `self.validation_docs()`,
            whichever is the main split used.
        :param repeats: int
        TODO: update this docstring
lintangsutawika's avatar
lintangsutawika committed
697
            The number of times each instance in a dataset is inferred on. Defaults to 1,
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
            can be increased for techniques like majority voting.
        """
        pass

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

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

Baber's avatar
Baber committed
715
    @deprecated("not used anymore")
716
717
718
719
720
721
722
723
    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
724
    @deprecated("not used anymore")
725
726
727
728
729
730
731
732
    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

733
734
735
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
736
737
738
739
740
741
742
743
744
745
    @classmethod
    def count_bytes(cls, doc):
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

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

746
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
747
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
748
749
750
751
752
753
754
        """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
755
756
757
758
759
        :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.
760
761
762
        :returns: str
            The fewshot context.
        """
763
        if rnd is None:
764
765
766
767
768
769
            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
770

771
        description = description if description else ""
772
773

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
774
            labeled_examples = ""
775
        else:
lintangsutawika's avatar
lintangsutawika committed
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
            # 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
800
            )
801
802

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
803
        return description + labeled_examples + example
804

805
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
806
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
807
808
        if hasattr(self, "_filters"):
            for f in self._filters:
809
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
810
811
812
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
813

baberabb's avatar
baberabb committed
814
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
815
        """Returns the config as a dictionary."""
816
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
817
        # (num_fewshot)
818
        return self.config.to_dict()
819

Baber Abbasi's avatar
Baber Abbasi committed
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
    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
842
843
844
845
846
847
848
        # 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
849

850
851
852
853
854
    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

855
856
857
858
859
860
861
    @property
    def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
862
863
864
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
865
866

    def doc_iterator(
867
868
869
870
871
872
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
873
    ) -> Iterator[Tuple[int, Any]]:
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
        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),
            )
896
897
        return doc_iterator

898
899

class ConfigurableTask(Task):
900
    VERSION = "Yaml"
901
    OUTPUT_TYPE = None
902
    CONFIG = None
903
904

    def __init__(
905
906
907
908
909
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Baber's avatar
Baber committed
910
    ) -> None:
911
        # Get pre-configured attributes
912
        self._config = self.CONFIG
913

914
        # Use new configurations if there was no preconfiguration
915
        if self.config is None:
916
            self._config = TaskConfig(**config)
917
918
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
919
            if config is not None:
920
                self._config.__dict__.update(config)
921

922
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
923
924
925
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
926

927
928
929
930
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

931
        if self.config.output_type is not None:
932
933
934
935
            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)}'"
                )
936
            self.OUTPUT_TYPE = self.config.output_type
937

938
939
940
941
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

942
943
944
945
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
946
947
948
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

949
950
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
951

952
953
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
954

Baber's avatar
Baber committed
955
        self.metric_list: list[MetricConfig] = self._config.get_metrics()
956

957
        self.download(self.config.dataset_kwargs)
958
959
960
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
961
962
        self._filters = self.config.get_filters()

963
964
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
965
            self.prompt = get_prompt(
966
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
967
            )
968
969
970
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
971
        if self.fewshot_docs() is not None:
972
973
974
975
            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
976
977
978
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
            )
            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)}"
                )
995

996
        self.task_docs = self.eval_docs
997

998
        # Test One Doc
999
        self.features = list(self.task_docs.features.keys())
1000
1001
        self.multiple_input = 0
        self.multiple_target = 0
1002
        test_doc = self.task_docs[0]
1003
        test_text = self.doc_to_text(test_doc)
1004
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
1005

1006
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1007
            test_choice = self.doc_to_choice(test_doc)
1008
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
1009
                eval_logger.error("doc_to_choice must return list")
1010
1011
            else:
                num_choice = len(test_choice)
1012

1013
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
1014
1015
1016
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
1017
                self.multiple_input = num_choice
1018
1019
        else:
            test_choice = None
1020

1021
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
1022
1023
1024
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
1025
            self.multiple_target = len(test_target)
1026
        else:
1027
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1028
                test_target = test_choice[test_target]
1029
            else:
lintangsutawika's avatar
lintangsutawika committed
1030
                test_target = str(test_target)
1031

1032
1033
1034
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1035
            check_choices = [test_target]
1036
1037
1038
1039
        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 = (
1040
1041
                    True
                    if self.config.target_delimiter.rstrip()
1042
                    != self.config.target_delimiter
1043
                    else False
1044
                )
1045

1046
                if delimiter_has_whitespace and choice_has_whitespace:
1047
1048
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1049
1050
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1051
                    eval_logger.debug(
1052
                        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'
1053
1054
                    )

Baber Abbasi's avatar
Baber Abbasi committed
1055
1056
1057
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
Baber Abbasi's avatar
Baber Abbasi committed
1058
1059
1060
1061
        from packaging.version import parse as vparse

        if dataset_kwargs and vparse(datasets.__version__) >= vparse("4.0.0"):
            dataset_kwargs.pop("trust_remote_code", None)
Baber Abbasi's avatar
Baber Abbasi committed
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
        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 {},
            )
1076

baberabb's avatar
baberabb committed
1077
    def has_training_docs(self) -> bool:
1078
        if self.config.training_split is not None:
1079
1080
1081
1082
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1083
    def has_validation_docs(self) -> bool:
1084
        if self.config.validation_split is not None:
1085
1086
1087
1088
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1089
    def has_test_docs(self) -> bool:
1090
        if self.config.test_split is not None:
1091
1092
1093
1094
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1095
    def training_docs(self) -> datasets.Dataset:
1096
        if self.has_training_docs():
1097
1098
1099
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1100
                )
1101
            return self.dataset[self.config.training_split]
1102

baberabb's avatar
baberabb committed
1103
    def validation_docs(self) -> datasets.Dataset:
1104
        if self.has_validation_docs():
1105
1106
1107
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1108
                )
1109
            return self.dataset[self.config.validation_split]
1110

baberabb's avatar
baberabb committed
1111
    def test_docs(self) -> datasets.Dataset:
1112
        if self.has_test_docs():
1113
1114
1115
            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]
1116

1117
    def fewshot_docs(self):
1118
        if self.config.fewshot_split is not None:
1119
1120
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1121
            return self.dataset[self.config.fewshot_split]
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
        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."
                )
1134
        else:
1135
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1136
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1137
                    f"[Task: {self.config.task}] "
1138
1139
1140
1141
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1142

KonradSzafer's avatar
KonradSzafer committed
1143
1144
1145
1146
1147
    @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
1148
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1149
1150
1151
1152
1153
1154
1155
1156
    ) -> 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":
1157
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1158
1159
            # if last message is user, append to it to avoid two user messages in a row
            else:
1160
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1161
1162
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1163
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1164
1165
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1166

lintangsutawika's avatar
lintangsutawika committed
1167
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1168
1169
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1170
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1171
1172
1173
1174
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1175
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1176
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1177
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1178
1179
1180
1181
1182
1183
1184
        """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
1185
1186
1187
1188
1189
1190
        :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.
1191
1192
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1193
1194
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1195
1196
1197
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1198
1199
1200
1201
1202
1203
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1204
1205
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1206

KonradSzafer's avatar
KonradSzafer committed
1207
1208
1209
1210
1211
1212
1213
1214
1215
        # 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
1216
        else:
KonradSzafer's avatar
KonradSzafer committed
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
            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
1230
1231
1232
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1233
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1234
1235
1236
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1237
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1238
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1239
                )
lintangsutawika's avatar
lintangsutawika committed
1240
1241

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1242
1243
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1244
                # TODO: append prefill?
1245
1246
                if not labeled_examples:
                    return ""
1247
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1248
1249
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1250
1251
1252
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1253
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1254
1255
1256
1257
1258
1259
1260
                )
            # 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
1261
1262
1263
1264
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1265
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1266
1267
1268
1269
1270
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1271
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1272
1273
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1274
1275
1276
1277
1278
1279
                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
1280
1281
1282
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1283
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1284
1285
1286
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1287
1288
1289
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1290
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1291
1292
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1293
1294
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1295
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1296
            )
1297
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1298
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1299
1300
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1301
1302
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1303
1304
            if self.multiple_input:
                return labeled_examples
1305
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1306
                return labeled_examples + example + prefix
1307
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1308
                return [labeled_examples + ex + prefix for ex in example]
1309
1310
1311
            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
1312
                    return labeled_examples + choices[example] + prefix
1313
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1314
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1315

Baber Abbasi's avatar
Baber Abbasi committed
1316
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1317
        """Iterates over FilterEnsembles and applies them to instances"""
1318
1319
        if hasattr(self, "_filters"):
            for f in self._filters:
1320
                f.apply(self._instances)
1321
1322
1323
1324
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1325
    def should_decontaminate(self):
1326
        return self.config.should_decontaminate
1327

Baber Abbasi's avatar
Baber Abbasi committed
1328
    def doc_to_decontamination_query(self, doc: dict):
1329
        if self.config.should_decontaminate:
1330
1331
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1332
            else:
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
                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
                        )
                    )
1344

1345
    def _process_doc(self, doc: dict) -> dict:
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
        """
        Override this to process (detokenize, strip, replace, etc.) individual
        documents. This can be used in a map over documents of a data split.
        E.g. `map(self._process_doc, self.dataset["validation"])`

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

Yu Shi Jie's avatar
Yu Shi Jie committed
1356
    def doc_to_text(self, doc, doc_to_text=None):
1357
1358
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1359
1360
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1361
        else:
1362
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1363

1364
        if isinstance(doc_to_text, int):
1365
            return doc_to_text
1366
        elif isinstance(doc_to_text, str):
1367
            if doc_to_text in self.features:
1368
                # if self.config.doc_to_choice is not None:
1369
1370
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1371
1372
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1373
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
Baber committed
1374
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1375
1376
1377
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1378
        elif callable(doc_to_text):
1379
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1380
        # Used when applying a Promptsource template
1381
        elif hasattr(doc_to_text, "apply"):
1382
1383
1384
1385
1386
            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")
1387
                return self.config.fewshot_delimiter
1388
        else:
1389
            print(type(doc_to_text))
1390
            raise TypeError
1391

Yu Shi Jie's avatar
Yu Shi Jie committed
1392
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1393
1394
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1395
1396
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1397
        else:
1398
            doc_to_target = self.config.doc_to_target
1399

1400
        if isinstance(doc_to_target, int):
1401
            return doc_to_target
1402
        elif isinstance(doc_to_target, str):
1403
            if doc_to_target in self.features:
1404
                # if self.config.doc_to_choice is not None:
1405
1406
1407
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1408
            else:
lintangsutawika's avatar
lintangsutawika committed
1409
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
Baber committed
1410
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1411
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1412
1413
1414
1415
1416
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1417
1418
1419
1420
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1421
1422
                else:
                    return target_string
1423
        elif isinstance(doc_to_target, list):
1424
            return doc_to_target
1425
        elif callable(doc_to_target):
1426
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1427
        # Used when applying a Promptsource template
1428
        elif hasattr(doc_to_target, "apply"):
1429
            applied_prompt = doc_to_target.apply(doc)
1430
1431
1432
1433
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1434
                return self.config.fewshot_delimiter
1435
1436
        else:
            raise TypeError
1437

Yu Shi Jie's avatar
Yu Shi Jie committed
1438
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1439
1440
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1441
1442
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1443
        elif self.config.doc_to_choice is None:
1444
1445
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1446
            doc_to_choice = self.config.doc_to_choice
1447

1448
        if isinstance(doc_to_choice, str):
1449
1450
1451
1452
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1453
        elif isinstance(doc_to_choice, list):
1454
            return doc_to_choice
1455
        elif isinstance(doc_to_choice, dict):
1456
1457
1458
1459
1460
1461
1462
            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
1463

1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list]:
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list]:
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber Abbasi's avatar
Baber Abbasi committed
1510
1511
1512
1513
1514
1515
1516
1517
    def doc_to_prefix(self, doc):
        if (gen_prefix := self.config.gen_prefix) is not None:
            if gen_prefix in self.features:
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1518
1519
1520
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1521
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1522
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1523

1524
1525
        aux_arguments = None

1526
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1527
            arguments = (ctx, self.doc_to_target(doc))
1528
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1529
            arguments = (self.doc_to_target(doc),)
1530
        elif self.OUTPUT_TYPE == "multiple_choice":
1531
            choices = self.doc_to_choice(doc)
1532
            target_delimiter = self.config.target_delimiter
1533
1534
            if apply_chat_template:
                target_delimiter = ""
1535
1536
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1537
                # apply chat_template to choices if apply_chat_template
1538
                cont = self.doc_to_target(doc)
1539

1540
                arguments = [
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1551
                ]
1552
            else:
1553
                # Otherwise they are placed in the continuation
1554
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1555

1556
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
Baber's avatar
Baber committed
1557
            if "acc_mutual_info" in [m.metric_names for m in self.metric_list]:
1558
1559
1560
1561
1562
1563
                # 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.
1564
1565
1566
1567
                # 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
                ]
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582

                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)},
            }

1583
1584
1585
1586
1587
1588
1589
1590
        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)},
            }

1591
1592
1593
1594
1595
1596
1597
        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":
1598
            request_list = [
1599
1600
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1601
                    doc=doc,
1602
                    arguments=arg,
1603
                    idx=i,
1604
1605
                    **kwargs,
                )
1606
                for i, arg in enumerate(arguments)
1607
            ]
1608
1609

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1610

lintangsutawika's avatar
lintangsutawika committed
1611
        return Instance(
1612
1613
1614
1615
1616
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1617
        )
1618
1619

    def process_results(self, doc, results):
1620
1621
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1622

1623
        result_dict = {}
Baber's avatar
Baber committed
1624
        use_metric = list(m.metric_names for m in self.metric_list)
1625
1626
1627
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1628
1629
1630
1631
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1632
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1633
            (loglikelihood,) = results
1634
1635
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1636
            return {
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
                **(
                    {"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
1652
            }
1653
        elif self.OUTPUT_TYPE == "multiple_choice":
1654
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1655

1656
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1657
            choices = self.doc_to_choice(doc)
1658
1659
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1660
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1661
1662
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1663
1664
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1665
1666
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1667
                # and this stores our "regular" conditional loglikelihoods
1668
                lls = lls[: len(choices)]
1669

1670
1671
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1672

1673
1674
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1675
            else:
1676
                gold = self.doc_to_target(doc)
1677
1678

            gold_index_error = False
1679
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1680
1681
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1682
1683
                    gold_index_error = True
            else:
1684
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1685
                    gold = gold if gold < len(choices) else -100
1686
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1687
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1688

Lintang Sutawika's avatar
Lintang Sutawika committed
1689
                if gold == -100:
1690
1691
1692
1693
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1694
                    f"Label index was not in within range of available choices,"
1695
1696
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1697

1698
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1699
1700
                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
1701
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1702
1703
1704
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1705
                # 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
1706
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1707

Lintang Sutawika's avatar
Lintang Sutawika committed
1708
1709
1710
1711
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1712
            result_dict = {
1713
                **({"acc": acc} if "acc" in use_metric else {}),
1714
1715
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1716
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1717
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1718
1719
1720
1721
1722
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1723
1724
            }

1725
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1726
1727
1728
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1729
1730
1731
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1732
        elif self.OUTPUT_TYPE == "generate_until":
1733
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1734
            result = results[0]
1735
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1736
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1737
                # it assumes that doc_to_target returns a number.
1738
1739
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1740
1741
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1742
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1743
1744
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
Baber's avatar
Baber committed
1745
                "bypass" in use_metric or isinstance(result, list)
1746
            ):
Chris's avatar
Chris committed
1747
1748
                # cast gold to the same type as result
                gold = type(result)(gold)
1749

Baber's avatar
Baber committed
1750
            for metric in self.metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
1751
1752
1753
1754
1755
                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
1756
1757
1758
1759
                    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
1760
                    if metric.name == "exact_match":
1761
                        result = [result for _ in range(len(gold))]
Baber's avatar
Baber committed
1762
                        scores = metric.fn(
1763
1764
                            references=gold,
                            predictions=result,
Baber's avatar
Baber committed
1765
                            **metric.kwargs,
1766
1767
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1768
                    else:
1769
1770
                        for gold_option in gold:
                            try:
Baber's avatar
Baber committed
1771
                                result_score = metric.fn(
1772
1773
                                    references=[gold_option],
                                    predictions=[result],
Baber's avatar
Baber committed
1774
                                    **metric.kwargs,
1775
1776
1777
1778
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
Baber's avatar
Baber committed
1779
                                result_score = metric.fn([gold_option, result])
1780
1781
1782
1783
1784
1785
1786
1787
                            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
1788
                else:
1789
                    try:
Baber's avatar
Baber committed
1790
                        result_score = metric.fn(
1791
1792
                            references=[gold],
                            predictions=[result],
Baber's avatar
Baber committed
1793
                            **metric.kwargs,
1794
                        )
1795
                    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
1796
                        result_score = metric.fn([gold, result])
1797
1798
1799
1800
1801
1802
1803
                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
1804
        else:
lintangsutawika's avatar
lintangsutawika committed
1805
1806
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1807
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1808
            )
1809
1810
1811

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1812
    def aggregation(self) -> dict:
Baber's avatar
Baber committed
1813
        return {k.name: k.aggregation_fn for k in self.metric_list}
1814

Baber Abbasi's avatar
Baber Abbasi committed
1815
    def higher_is_better(self) -> dict:
Baber's avatar
Baber committed
1816
        return {k.name: k.higher_is_better for k in self.metric_list}
1817

Baber Abbasi's avatar
Baber Abbasi committed
1818
1819
1820
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1821
1822
1823
1824
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1825
1826
1827
1828
1829
    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
1830
            f"num_samples={len(self.eval_docs)})"
1831
1832
        )

1833
1834

class MultipleChoiceTask(Task):
1835
    OUTPUT_TYPE = "loglikelihood"
1836

baberabb's avatar
baberabb committed
1837
    def doc_to_target(self, doc: dict) -> str:
1838
1839
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1840
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1841
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1842
1843
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1844
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1845
                doc=doc,
1846
                arguments=(ctx, " {}".format(choice)),
1847
                idx=i,
1848
1849
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1850
1851
            for i, choice in enumerate(doc["choices"])
        ]
1852

1853
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1854
1855
1856
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1857
1858
1859
1860
1861
1862
1863
1864
1865
1866
1867
        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
1868
    def higher_is_better(self) -> dict:
1869
1870
1871
1872
1873
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1874
    def aggregation(self) -> dict:
1875
1876
1877
1878
1879
1880
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1881
class PerplexityTask(Task):
1882
1883
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1884
    def has_training_docs(self) -> bool:
1885
1886
        return False

baberabb's avatar
baberabb committed
1887
    def fewshot_examples(self, k: int, rnd) -> List:
1888
1889
1890
1891
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1892
1893
        return []

baberabb's avatar
baberabb committed
1894
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1895
1896
1897
1898
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1899
1900
1901

        return ""

baberabb's avatar
baberabb committed
1902
    def higher_is_better(self) -> dict:
1903
1904
1905
1906
1907
1908
1909
1910
1911
        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
1912
    def doc_to_text(self, doc) -> str:
1913
1914
1915
1916
1917
        return ""

    def doc_to_target(self, doc):
        return doc

1918
1919
1920
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1921

lintangsutawika's avatar
lintangsutawika committed
1922
1923
1924
1925
1926
1927
1928
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1929

1930
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1931
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1932
1933
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1934
1935
1936
1937
1938
1939
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1940
    def aggregation(self) -> dict:
1941
1942
1943
1944
1945
1946
1947
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1948
    def count_bytes(cls, doc) -> int:
1949
1950
1951
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1952
    def count_words(cls, doc) -> int:
1953
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1954
        return len(re.split(r"\s+", doc))