task.py 74.6 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
    is_elementwise: bool = True
Baber's avatar
Baber committed
67
68

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

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

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

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

Baber's avatar
Baber committed
90

91
92
93
94
95
96
97
98
99
@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
100
101
@dataclass
class FilterConfig:
Baber's avatar
nit  
Baber committed
102
    """Encapsulates information about a single filter."""
Baber's avatar
Baber committed
103
104
105
106
107
108

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


Baber's avatar
Baber committed
109
110
111
112
@dataclass
class FewshotConfig:
    sampler: str
    samples: list[dict]
113
    process_docs: Optional[Callable] = None
Baber's avatar
Baber committed
114
115


Baber's avatar
Baber committed
116
117
118
119
120
121
122
123
124
125
@dataclass
class DatasetConfig:
    """Encapsulates information about a dataset."""

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


126
127
@dataclass
class TaskConfig(dict):
128
    # task naming/registry
129
130
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
131
    tag: Optional[Union[str, list]] = None
132
133
134
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
135
    custom_dataset: Optional[Callable] = None
136
137
138
139
140
141
    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
142
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
143
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
144
    )
145
146
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
147
148
149
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
Baber's avatar
Baber committed
150
151
    doc_to_image: Union[Callable, str, None] = None
    doc_to_audio: Union[Callable, str, None] = None
Hojin Lee's avatar
Hojin Lee committed
152
    unsafe_code: bool = False
153
154
155
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
156
    description: str = ""
157
158
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
159
    fewshot_config: Optional[dict] = None
160
    # runtime configuration options
161
    num_fewshot: Optional[int] = None
162
    # scoring options
163
164
165
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
166
    repeats: int = 1
167
    filter_list: Optional[Union[str, list]] = None
168
    should_decontaminate: bool = False
169
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
170
    gen_prefix: Optional[str] = None
171
172
173
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
Baber's avatar
nit  
Baber committed
174
175
    _metric_list = None
    _filter_list = None
176

Ethan Smith's avatar
Ethan Smith committed
177
    def __post_init__(self) -> None:
Lintang Sutawika's avatar
Lintang Sutawika committed
178
        if self.generation_kwargs is not None:
179
            if self.output_type != "generate_until":
180
                eval_logger.warning(
181
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
Lintang Sutawika's avatar
Lintang Sutawika committed
182
183
184
185
186
187
188
189
                )

            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
190
191
192
                eval_logger.warning(
                    f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                )
193
                self.generation_kwargs["until"] = [self.fewshot_delimiter]
Lintang Sutawika's avatar
Lintang Sutawika committed
194
        else:
195
            if self.output_type == "generate_until":
Lintang Sutawika's avatar
Lintang Sutawika committed
196
197
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
198
199
200
201
202
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
203
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
204
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
205
                }
Baber Abbasi's avatar
Baber Abbasi committed
206
207
208
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
209

Baber's avatar
Baber committed
210
211
212
213
214
215
216
217
218
219
220
        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
221
222
223
            eval_logger.info(
                f"No metrics defined in config, using default metrics for {self.output_type}={_metric_list}"
            )
Baber's avatar
Baber committed
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
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
            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
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
    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

329
330
331
    def __getitem__(self, item):
        return getattr(self, item)

332
333
334
    def __setitem__(self, item, value):
        return setattr(self, item, value)

335
    def to_dict(self, keep_callable: bool = False) -> dict:
336
337
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
338
        Used for dumping results alongside full task configuration
339

haileyschoelkopf's avatar
haileyschoelkopf committed
340
341
342
343
344
345
346
347
348
349
        :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)
350
351
352
353
354
355
356
357
358
359
            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
360
        return cfg_dict
361

362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
    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)

378
379
380
381
382
383
384
385
386
387
388

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)
    """

389
    VERSION: Optional[Union[int, str]] = None
390

391
392
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
393
    DATASET_PATH: Optional[str] = None
394
395

    # The name of a subset within `DATASET_PATH`.
396
    DATASET_NAME: Optional[str] = None
397

398
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
399

400
401
    def __init__(
        self,
402
403
404
405
        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
406
    ) -> None:
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
        """
        :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)
429
430
431
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
432

433
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
434

lintangsutawika's avatar
lintangsutawika committed
435
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
436
437
438
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
439

440
441
442
443
444
445
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
        """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.
        """
470
471
472
473
474
475
476
        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,
        )
477

478
    @property
479
    def config(self) -> TaskConfig:
480
481
482
        """Returns the TaskConfig associated with this class."""
        return self._config

483
    @abc.abstractmethod
Baber's avatar
Baber committed
484
    def has_training_docs(self) -> bool:
485
486
487
488
        """Whether the task has a training set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
489
    def has_validation_docs(self) -> bool:
490
491
492
493
        """Whether the task has a validation set"""
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
494
    def has_test_docs(self) -> bool:
495
496
497
        """Whether the task has a test set"""
        pass

498
    def training_docs(self) -> Iterable:
499
500
501
502
503
504
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

505
    def validation_docs(self) -> Iterable:
506
507
508
509
510
511
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

512
    def test_docs(self) -> Iterable:
513
514
515
516
517
518
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

519
    def fewshot_docs(self) -> Iterable:
520
521
522
523
524
525
526
527
528
        """
        :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
529
530
531
532
533
            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."
                )
534
535
            return self.test_docs()

536
    def _process_doc(self, doc: dict) -> dict:
537
538
539
540
541
542
543
544
545
        """
        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
546

547
    @property
548
    def instances(self) -> List[Instance]:
549
550
551
552
553
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

Baber's avatar
Baber committed
554
    def fewshot_examples(self, k, rnd) -> Iterable[dict]:
555
556
557
558
559
        if self._training_docs is None:
            self._training_docs = list(self.training_docs())

        return rnd.sample(self._training_docs, k)

560
561
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
562
563
564
565
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

    @abc.abstractmethod
Baber's avatar
Baber committed
566
    def doc_to_text(self, doc) -> str:
567
568
569
        pass

    @abc.abstractmethod
Baber's avatar
Baber committed
570
    def doc_to_target(self, doc) -> Union[str, int]:
571
572
        pass

573
574
575
576
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

577
578
579
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber's avatar
Baber committed
580
    def doc_to_prefix(self, doc) -> str:
Baber Abbasi's avatar
Baber Abbasi committed
581
582
        return ""

583
584
    def build_all_requests(
        self,
585
        *,
586
        limit: Union[int, None] = None,
587
        samples: Optional[List[int]] = None,
588
589
590
591
592
593
594
595
596
        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 = "",
597
    ) -> None:
598
        """Build a set of Instances for a task, and store them in task.instances"""
599
600
601
602

        # used with caching
        og_limit = limit

603
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
604
605
606
607
608
609
610
        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 ""
        )
611
        cache_key += f"-tokenizer{tokenizer_name}"
612

Baber Abbasi's avatar
Baber Abbasi committed
613
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
614
615
616
617
618
619
620
621
622
623
624
625
626

        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
627
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
628

629
        instances = []
630
631
632
633
634
635
636
637
638
639

        # 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(
640
641
642
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
643
644
645
646
647
648
649
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
650
        ):
651
            # sample fewshot context #TODO: need to offset doc_id by rank now!
652
            fewshot_ctx = self.fewshot_context(
653
                doc,
654
655
656
657
658
659
660
                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
661
                gen_prefix=self.doc_to_prefix(doc),
662
            )
663

664
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
665
666
667
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
668
                metadata=(self.config["task"], doc_id, self.config.repeats),
669
                apply_chat_template=apply_chat_template,
670
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
671
            )
672
673
674
675

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

676
677
678
679
680
681
682
683
684
685
686
687
688
            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
689

690
691
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
692

693
694
695
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
    @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
712
            The number of times each instance in a dataset is inferred on. Defaults to 1,
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
            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
730
    @deprecated("not used anymore")
731
732
733
734
735
736
737
738
    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
739
    @deprecated("not used anymore")
740
741
742
743
744
745
746
747
    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

748
749
750
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
751
    @classmethod
Baber's avatar
Baber committed
752
    def count_bytes(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
753
754
755
756
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

    @classmethod
Baber's avatar
Baber committed
757
    def count_words(cls, doc) -> int:
haileyschoelkopf's avatar
haileyschoelkopf committed
758
759
760
        """Downstream loglikelihood_rolling perplexity tasks with custom word boundaries should override this!"""
        return len(re.split(r"\s+", doc))

761
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
762
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
763
764
765
766
767
768
769
        """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
770
771
772
773
774
        :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.
775
776
777
        :returns: str
            The fewshot context.
        """
778
        if rnd is None:
779
780
781
782
783
784
            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
785

786
        description = description if description else ""
787
788

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
789
            labeled_examples = ""
790
        else:
lintangsutawika's avatar
lintangsutawika committed
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
            # 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
815
            )
816
817

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
818
        return description + labeled_examples + example
819

820
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
821
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
822
823
        if hasattr(self, "_filters"):
            for f in self._filters:
824
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
825
826
827
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
828

baberabb's avatar
baberabb committed
829
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
830
        """Returns the config as a dictionary."""
831
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
832
        # (num_fewshot)
833
        return self.config.to_dict()
834

Baber Abbasi's avatar
Baber Abbasi committed
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
    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
857
858
859
860
861
862
863
        # 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
864

865
866
867
868
869
    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

870
    @property
Baber's avatar
Baber committed
871
    def eval_docs(self) -> Union[datasets.Dataset, Iterable[dict]]:
872
873
874
875
876
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
877
878
879
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
880
881

    def doc_iterator(
882
883
884
885
886
887
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
888
    ) -> Iterator[Tuple[int, Any]]:
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
        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),
            )
911
912
        return doc_iterator

913
914

class ConfigurableTask(Task):
915
    VERSION = "Yaml"
916
    OUTPUT_TYPE = None
917
    CONFIG = None
918
919

    def __init__(
920
921
922
923
924
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Baber's avatar
Baber committed
925
    ) -> None:
926
        # Get pre-configured attributes
927
        self._config = self.CONFIG
928

929
        # Use new configurations if there was no preconfiguration
930
        if self.config is None:
931
            self._config = TaskConfig(**config)
932
933
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
934
            if config is not None:
935
                self._config.__dict__.update(config)
936

937
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
938
939
940
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
941

942
943
944
945
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

946
        if self.config.output_type is not None:
947
948
949
950
            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)}'"
                )
951
            self.OUTPUT_TYPE = self.config.output_type
952

953
954
955
956
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

957
958
959
960
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
961
962
963
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

964
965
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
966

967
968
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
969

Baber's avatar
Baber committed
970
        self.metric_list: list[MetricConfig] = self.config.get_metrics()
971

972
        self.download(self.config.dataset_kwargs)
973
974
975
        self._training_docs = None
        self._fewshot_docs = None

Baber's avatar
Baber committed
976
977
        self._filters = self.config.get_filters()

978
979
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
980
            self.prompt = get_prompt(
981
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
982
            )
983
984
985
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
986
        if self.fewshot_docs() is not None:
987
988
989
990
            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
991
992
993
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
            )
            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)}"
                )
1010

1011
        self.task_docs = self.eval_docs
1012

1013
        # Test One Doc
1014
        self.features = list(self.task_docs.features.keys())
1015
1016
        self.multiple_input = 0
        self.multiple_target = 0
1017
        test_doc = self.task_docs[0]
1018
        test_text = self.doc_to_text(test_doc)
1019
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
1020

1021
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1022
            test_choice = self.doc_to_choice(test_doc)
1023
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
1024
                eval_logger.error("doc_to_choice must return list")
1025
1026
            else:
                num_choice = len(test_choice)
1027

1028
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
1029
1030
1031
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
1032
                self.multiple_input = num_choice
1033
1034
        else:
            test_choice = None
1035

1036
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
1037
1038
1039
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
1040
            self.multiple_target = len(test_target)
1041
        else:
1042
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
1043
                test_target = test_choice[test_target]
1044
            else:
lintangsutawika's avatar
lintangsutawika committed
1045
                test_target = str(test_target)
1046

1047
1048
1049
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
1050
            check_choices = [test_target]
1051
1052
1053
1054
        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 = (
1055
1056
                    True
                    if self.config.target_delimiter.rstrip()
1057
                    != self.config.target_delimiter
1058
                    else False
1059
                )
1060

1061
                if delimiter_has_whitespace and choice_has_whitespace:
1062
1063
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
1064
1065
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
1066
                    eval_logger.debug(
1067
                        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'
1068
1069
                    )

Baber Abbasi's avatar
Baber Abbasi committed
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
    def download(
        self, dataset_kwargs: Optional[Dict[str, Any]] = None, **kwargs
    ) -> None:
        if isinstance(self.config.custom_dataset, Callable):
            eval_logger.warning(
                f"{self.config.task}: Custom kwargs can be passed to `--metadata` in console (as json string) or to the TaskManager."
                + "\nFor example --metadata='{\"max_seq_lengths\":[4096, 8192]}'. For details see task Readme."
            )
            self.dataset = self.config.custom_dataset(
                **(self.config.metadata or {}), **(self.config.dataset_kwargs or {})
            )
        else:
            self.dataset = datasets.load_dataset(
                path=self.DATASET_PATH,
                name=self.DATASET_NAME,
                **dataset_kwargs if dataset_kwargs is not None else {},
            )
1087

baberabb's avatar
baberabb committed
1088
    def has_training_docs(self) -> bool:
1089
        if self.config.training_split is not None:
1090
1091
1092
1093
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1094
    def has_validation_docs(self) -> bool:
1095
        if self.config.validation_split is not None:
1096
1097
1098
1099
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1100
    def has_test_docs(self) -> bool:
1101
        if self.config.test_split is not None:
1102
1103
1104
1105
            return True
        else:
            return False

Baber's avatar
Baber committed
1106
    def training_docs(self) -> Optional[datasets.Dataset]:
1107
        if self.has_training_docs():
1108
1109
1110
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1111
                )
1112
            return self.dataset[self.config.training_split]
1113

Baber's avatar
Baber committed
1114
    def validation_docs(self) -> Optional[datasets.Dataset]:
1115
        if self.has_validation_docs():
1116
1117
1118
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1119
                )
1120
            return self.dataset[self.config.validation_split]
1121

Baber's avatar
Baber committed
1122
    def test_docs(self) -> Optional[datasets.Dataset]:
1123
        if self.has_test_docs():
1124
1125
1126
            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]
1127

1128
    def fewshot_docs(self):
1129
        if self.config.fewshot_split is not None:
1130
1131
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1132
            return self.dataset[self.config.fewshot_split]
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
        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."
                )
1145
        else:
1146
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1147
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1148
                    f"[Task: {self.config.task}] "
1149
1150
1151
1152
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1153

KonradSzafer's avatar
KonradSzafer committed
1154
1155
1156
1157
1158
    @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
1159
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1160
1161
1162
1163
1164
1165
1166
1167
    ) -> 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":
1168
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1169
1170
            # if last message is user, append to it to avoid two user messages in a row
            else:
1171
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1172
1173
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1174
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1175
1176
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1177

lintangsutawika's avatar
lintangsutawika committed
1178
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1179
1180
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1181
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1182
1183
1184
1185
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1186
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1187
        gen_prefix: Optional[str] = None,
Baber's avatar
Baber committed
1188
    ) -> Union[str, List[str], None]:
lintangsutawika's avatar
lintangsutawika committed
1189
1190
1191
1192
1193
1194
1195
        """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
1196
1197
1198
1199
1200
1201
        :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.
1202
1203
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1204
1205
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1206
1207
1208
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1209
1210
1211
1212
1213
1214
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1215
1216
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1217

KonradSzafer's avatar
KonradSzafer committed
1218
1219
1220
1221
1222
1223
1224
1225
1226
        # 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
1227
        else:
KonradSzafer's avatar
KonradSzafer committed
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
            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
1241
1242
1243
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1244
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1245
1246
1247
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1248
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1249
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1250
                )
lintangsutawika's avatar
lintangsutawika committed
1251
1252

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1253
1254
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1255
                # TODO: append prefill?
1256
1257
                if not labeled_examples:
                    return ""
1258
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1259
1260
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1261
1262
1263
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1264
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1265
1266
1267
1268
1269
1270
1271
                )
            # 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
1272
1273
1274
1275
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1276
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1277
1278
1279
1280
1281
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1282
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1283
1284
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1285
1286
1287
1288
1289
1290
                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
1291
1292
1293
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1294
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1295
1296
1297
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1298
1299
1300
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1301
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1302
1303
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1304
1305
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1306
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1307
            )
1308
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1309
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1310
1311
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1312
1313
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1314
1315
            if self.multiple_input:
                return labeled_examples
1316
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1317
                return labeled_examples + example + prefix
1318
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1319
                return [labeled_examples + ex + prefix for ex in example]
1320
1321
1322
            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
1323
                    return labeled_examples + choices[example] + prefix
1324
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1325
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1326

Baber Abbasi's avatar
Baber Abbasi committed
1327
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1328
        """Iterates over FilterEnsembles and applies them to instances"""
1329
1330
        if hasattr(self, "_filters"):
            for f in self._filters:
1331
                f.apply(self._instances)
1332
1333
1334
1335
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1336
    def should_decontaminate(self):
1337
        return self.config.should_decontaminate
1338

Baber Abbasi's avatar
Baber Abbasi committed
1339
    def doc_to_decontamination_query(self, doc: dict):
1340
        if self.config.should_decontaminate:
1341
1342
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1343
            else:
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
                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
                        )
                    )
1355

1356
    def _process_doc(self, doc: dict) -> dict:
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
        """
        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
1367
    def doc_to_text(self, doc, doc_to_text=None):
1368
1369
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1370
1371
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1372
        else:
1373
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1374

1375
        if isinstance(doc_to_text, int):
1376
            return doc_to_text
1377
        elif isinstance(doc_to_text, str):
1378
            if doc_to_text in self.features:
1379
                # if self.config.doc_to_choice is not None:
1380
1381
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1382
1383
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1384
                text_string = utils.apply_template(doc_to_text, doc)
Baber's avatar
nit  
Baber committed
1385
                if text_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1386
1387
1388
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1389
        elif callable(doc_to_text):
1390
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1391
        # Used when applying a Promptsource template
1392
        elif hasattr(doc_to_text, "apply"):
1393
1394
1395
1396
1397
            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")
1398
                return self.config.fewshot_delimiter
1399
        else:
1400
            print(type(doc_to_text))
1401
            raise TypeError
1402

Yu Shi Jie's avatar
Yu Shi Jie committed
1403
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1404
1405
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1406
1407
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1408
        else:
1409
            doc_to_target = self.config.doc_to_target
1410

1411
        if isinstance(doc_to_target, int):
1412
            return doc_to_target
1413
        elif isinstance(doc_to_target, str):
1414
            if doc_to_target in self.features:
1415
                # if self.config.doc_to_choice is not None:
1416
1417
1418
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1419
            else:
lintangsutawika's avatar
lintangsutawika committed
1420
                target_string = utils.apply_template(doc_to_target, doc)
Baber's avatar
nit  
Baber committed
1421
                if target_string.isdigit() and self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1422
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1423
1424
1425
1426
1427
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1428
1429
1430
1431
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1432
1433
                else:
                    return target_string
1434
        elif isinstance(doc_to_target, list):
1435
            return doc_to_target
1436
        elif callable(doc_to_target):
1437
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1438
        # Used when applying a Promptsource template
1439
        elif hasattr(doc_to_target, "apply"):
1440
            applied_prompt = doc_to_target.apply(doc)
1441
1442
1443
1444
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1445
                return self.config.fewshot_delimiter
1446
1447
        else:
            raise TypeError
1448

Yu Shi Jie's avatar
Yu Shi Jie committed
1449
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1450
1451
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1452
1453
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1454
        elif self.config.doc_to_choice is None:
1455
1456
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1457
            doc_to_choice = self.config.doc_to_choice
1458

1459
        if isinstance(doc_to_choice, str):
1460
1461
1462
1463
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1464
        elif isinstance(doc_to_choice, list):
1465
            return doc_to_choice
1466
        elif isinstance(doc_to_choice, dict):
1467
1468
1469
1470
1471
1472
1473
            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
1474

Baber's avatar
Baber committed
1475
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list, None]:
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

Baber's avatar
Baber committed
1498
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list, None]:
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber's avatar
Baber committed
1521
    def doc_to_prefix(self, doc) -> Optional[str]:
Baber Abbasi's avatar
Baber Abbasi committed
1522
1523
1524
1525
1526
1527
1528
        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
1529
1530
1531
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1532
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1533
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1534

1535
1536
        aux_arguments = None

1537
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1538
            arguments = (ctx, self.doc_to_target(doc))
1539
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1540
            arguments = (self.doc_to_target(doc),)
1541
        elif self.OUTPUT_TYPE == "multiple_choice":
1542
            choices = self.doc_to_choice(doc)
1543
            target_delimiter = self.config.target_delimiter
1544
1545
            if apply_chat_template:
                target_delimiter = ""
1546
1547
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1548
                # apply chat_template to choices if apply_chat_template
1549
                cont = self.doc_to_target(doc)
1550

1551
                arguments = [
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1562
                ]
1563
            else:
1564
                # Otherwise they are placed in the continuation
1565
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1566

1567
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
Baber's avatar
Baber committed
1568
            if "acc_mutual_info" in [m.metric_name for m in self.metric_list]:
1569
1570
1571
1572
1573
1574
                # 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.
1575
1576
1577
1578
                # 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
                ]
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593

                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)},
            }

1594
1595
1596
1597
1598
1599
1600
1601
        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)},
            }

1602
1603
1604
1605
1606
1607
1608
        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":
1609
            request_list = [
1610
1611
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1612
                    doc=doc,
1613
                    arguments=arg,
1614
                    idx=i,
1615
1616
                    **kwargs,
                )
1617
                for i, arg in enumerate(arguments)
1618
            ]
1619
1620

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1621

lintangsutawika's avatar
lintangsutawika committed
1622
        return Instance(
1623
1624
1625
1626
1627
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=arguments,
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1628
        )
1629
1630

    def process_results(self, doc, results):
1631
1632
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1633

1634
        result_dict = {}
Baber's avatar
Baber committed
1635
        use_metric = list(m.metric_name for m in self.metric_list)
1636
1637
1638
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1639
1640
1641
1642
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1643
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1644
            (loglikelihood,) = results
1645
1646
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1647
            return {
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
                **(
                    {"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
1663
            }
1664
        elif self.OUTPUT_TYPE == "multiple_choice":
1665
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1666

1667
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1668
            choices = self.doc_to_choice(doc)
1669
1670
            completion_len = np.array([float(len(i)) for i in choices])

Baber's avatar
Baber committed
1671
            if 2 * len(choices) == len(lls) and "acc_mutual_info" in use_metric:
1672
1673
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1674
1675
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1676
1677
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1678
                # and this stores our "regular" conditional loglikelihoods
1679
                lls = lls[: len(choices)]
1680

1681
1682
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1683

1684
1685
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1686
            else:
1687
                gold = self.doc_to_target(doc)
1688
1689

            gold_index_error = False
1690
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1691
1692
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1693
1694
                    gold_index_error = True
            else:
1695
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1696
                    gold = gold if gold < len(choices) else -100
1697
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1698
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1699

Lintang Sutawika's avatar
Lintang Sutawika committed
1700
                if gold == -100:
1701
1702
1703
1704
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1705
                    f"Label index was not in within range of available choices,"
1706
1707
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1708

1709
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1710
1711
                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
1712
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1713
1714
1715
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1716
                # 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
1717
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1718

Lintang Sutawika's avatar
Lintang Sutawika committed
1719
1720
1721
1722
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1723
            result_dict = {
1724
                **({"acc": acc} if "acc" in use_metric else {}),
1725
1726
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1727
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1728
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1729
1730
1731
1732
1733
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1734
1735
            }

1736
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1737
1738
1739
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1740
1741
1742
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1743
        elif self.OUTPUT_TYPE == "generate_until":
1744
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1745
            result = results[0]
1746
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1747
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1748
                # it assumes that doc_to_target returns a number.
1749
1750
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1751
1752
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1753
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1754
1755
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
Baber's avatar
Baber committed
1756
                "bypass" in use_metric or isinstance(result, list)
1757
            ):
Chris's avatar
Chris committed
1758
1759
                # cast gold to the same type as result
                gold = type(result)(gold)
1760

Baber's avatar
Baber committed
1761
            for metric in self.metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
1762
1763
1764
1765
1766
                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
1767
1768
1769
1770
                    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
1771
                    if metric.name == "exact_match":
1772
                        result = [result for _ in range(len(gold))]
Baber's avatar
Baber committed
1773
                        scores = metric.fn(
1774
1775
                            references=gold,
                            predictions=result,
Baber's avatar
Baber committed
1776
                            **metric.kwargs,
1777
1778
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1779
                    else:
1780
1781
                        for gold_option in gold:
                            try:
Baber's avatar
Baber committed
1782
                                result_score = metric.fn(
1783
1784
                                    references=[gold_option],
                                    predictions=[result],
Baber's avatar
Baber committed
1785
                                    **metric.kwargs,
1786
1787
1788
1789
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
Baber's avatar
Baber committed
1790
                                result_score = metric.fn([gold_option, result])
1791
1792
1793
1794
1795
1796
1797
1798
                            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
1799
                else:
1800
                    try:
Baber's avatar
Baber committed
1801
                        result_score = metric.fn(
1802
1803
                            references=[gold],
                            predictions=[result],
Baber's avatar
Baber committed
1804
                            **metric.kwargs,
1805
                        )
1806
                    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
1807
                        result_score = metric.fn([gold, result])
1808
1809
1810
1811
1812
1813
1814
                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
1815
        else:
lintangsutawika's avatar
lintangsutawika committed
1816
1817
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1818
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1819
            )
1820
1821
1822

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1823
    def aggregation(self) -> dict:
Baber's avatar
Baber committed
1824
        return {k.name: k.aggregation_fn for k in self.metric_list}
1825

Baber Abbasi's avatar
Baber Abbasi committed
1826
    def higher_is_better(self) -> dict:
Baber's avatar
Baber committed
1827
        return {k.name: k.higher_is_better for k in self.metric_list}
1828

Baber Abbasi's avatar
Baber Abbasi committed
1829
1830
1831
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1832
    @property
Baber's avatar
Baber committed
1833
    def task_name(self) -> Optional[str]:
Lintang Sutawika's avatar
Lintang Sutawika committed
1834
1835
        return getattr(self.config, "task", None)

1836
1837
1838
1839
1840
    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
1841
            f"num_samples={len(self.eval_docs)})"
1842
1843
        )

1844
1845

class MultipleChoiceTask(Task):
1846
    OUTPUT_TYPE = "loglikelihood"
1847

baberabb's avatar
baberabb committed
1848
    def doc_to_target(self, doc: dict) -> str:
1849
1850
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1851
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1852
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1853
1854
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1855
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1856
                doc=doc,
1857
                arguments=(ctx, " {}".format(choice)),
1858
                idx=i,
1859
1860
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1861
1862
            for i, choice in enumerate(doc["choices"])
        ]
1863

1864
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1865
1866
1867
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1868
1869
1870
1871
1872
1873
1874
1875
1876
1877
1878
        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
1879
    def higher_is_better(self) -> dict:
1880
1881
1882
1883
1884
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1885
    def aggregation(self) -> dict:
1886
1887
1888
1889
1890
1891
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1892
class PerplexityTask(Task):
1893
1894
    OUTPUT_TYPE = "loglikelihood_rolling"

baberabb's avatar
baberabb committed
1895
    def has_training_docs(self) -> bool:
1896
1897
        return False

baberabb's avatar
baberabb committed
1898
    def fewshot_examples(self, k: int, rnd) -> List:
1899
1900
1901
1902
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1903
1904
        return []

baberabb's avatar
baberabb committed
1905
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1906
1907
1908
1909
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1910
1911
1912

        return ""

baberabb's avatar
baberabb committed
1913
    def higher_is_better(self) -> dict:
1914
1915
1916
1917
1918
1919
1920
1921
1922
        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
1923
    def doc_to_text(self, doc) -> str:
1924
1925
1926
1927
1928
        return ""

    def doc_to_target(self, doc):
        return doc

1929
1930
1931
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
1932

lintangsutawika's avatar
lintangsutawika committed
1933
1934
1935
1936
1937
1938
1939
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
1940

1941
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
1942
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
1943
1944
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
1945
1946
1947
1948
1949
1950
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
1951
    def aggregation(self) -> dict:
1952
1953
1954
1955
1956
1957
1958
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
1959
    def count_bytes(cls, doc) -> int:
1960
1961
1962
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
1963
    def count_words(cls, doc) -> int:
1964
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
1965
        return len(re.split(r"\s+", doc))