task.py 78.7 KB
Newer Older
1
import abc
2
import ast
Baber's avatar
Baber committed
3
import itertools
lintangsutawika's avatar
lintangsutawika committed
4
import logging
5
import random
6
import re
Baber's avatar
test  
Baber committed
7
from collections import defaultdict
8
from collections.abc import Callable
9
from copy import deepcopy
10
from dataclasses import asdict, dataclass
11
from inspect import getsource
12
13
14
15
16
17
18
19
20
21
22
23
from typing import (
    Any,
    Dict,
    Iterable,
    Iterator,
    List,
    Literal,
    Mapping,
    Optional,
    Tuple,
    Union,
)
24
25
26

import datasets
import numpy as np
27
from tqdm import tqdm
28
29

from lm_eval import utils
30
from lm_eval.api import samplers
31
from lm_eval.api.instance import Instance, OutputType
Baber's avatar
Baber committed
32
33
34
35
36
37
from lm_eval.api.metrics import (
    bits_per_byte,
    mean,
    stderr_for_metric,
    weighted_perplexity,
)
lintangsutawika's avatar
lintangsutawika committed
38
from lm_eval.api.registry import (
39
40
    AGGREGATION_REGISTRY,
    DEFAULT_METRIC_REGISTRY,
haileyschoelkopf's avatar
haileyschoelkopf committed
41
    get_aggregation,
42
    get_metric,
43
    get_metric_aggregation,
haileyschoelkopf's avatar
haileyschoelkopf committed
44
    is_higher_better,
lintangsutawika's avatar
lintangsutawika committed
45
)
46
from lm_eval.caching.cache import load_from_cache, save_to_cache
47
48
from lm_eval.filters import build_filter_ensemble
from lm_eval.prompts import get_prompt
Baber's avatar
Baber committed
49
from lm_eval.utils import create_sample_log, pass_at_k
50

51

52
53
54
55
ALL_OUTPUT_TYPES = [
    "loglikelihood",
    "multiple_choice",
    "loglikelihood_rolling",
56
    "generate_until",
57
58
]

Lintang Sutawika's avatar
Lintang Sutawika committed
59
eval_logger = logging.getLogger(__name__)
60

lintangsutawika's avatar
lintangsutawika committed
61

62
63
@dataclass
class TaskConfig(dict):
64
    # task naming/registry
65
66
    task: Optional[str] = None
    task_alias: Optional[str] = None
Lintang Sutawika's avatar
Lintang Sutawika committed
67
    tag: Optional[Union[str, list]] = None
68
69
70
    # HF dataset options.
    # which dataset to use,
    # and what splits for what purpose
Baber Abbasi's avatar
Baber Abbasi committed
71
    custom_dataset: Optional[Callable] = None
72
73
74
75
76
77
    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
78
    fewshot_split: Optional[str] = (
Baber Abbasi's avatar
Baber Abbasi committed
79
        None  # TODO: assert that this not None if num_fewshot > 0. (?) assert if this is same split as one evaluating (?)
80
    )
81
82
    # formatting / prompting options.
    # see docs/advanced_task_guide.md for more info
83
84
85
    process_docs: Optional[Callable] = None
    doc_to_text: Optional[Union[Callable, str]] = None
    doc_to_target: Optional[Union[Callable, str]] = None
86
    doc_to_image: Union[Callable, str] = None
87
    doc_to_audio: Union[Callable, str] = None
Hojin Lee's avatar
Hojin Lee committed
88
    unsafe_code: bool = False
89
90
91
    doc_to_choice: Optional[Union[Callable, str, dict, list]] = None
    process_results: Optional[Union[Callable, str]] = None
    use_prompt: Optional[str] = None
92
    description: str = ""
93
94
    target_delimiter: str = " "
    fewshot_delimiter: str = "\n\n"
95
    fewshot_config: Optional[dict] = None
96
    # runtime configuration options
97
    num_fewshot: Optional[int] = None
98
    # scoring options
99
100
101
    metric_list: Optional[list] = None
    output_type: OutputType = "generate_until"
    generation_kwargs: Optional[dict] = None
102
    repeats: int = 1
103
    filter_list: Optional[Union[str, list]] = None
104
    should_decontaminate: bool = False
105
    doc_to_decontamination_query: Optional[str] = None
Baber Abbasi's avatar
Baber Abbasi committed
106
    gen_prefix: Optional[str] = None
107
    repeat_agg: Optional[str] = None
108
109
110
    metadata: Optional[dict] = (
        None  # by default, not used in the code. allows for users to pass arbitrary info to tasks
    )
111

Ethan Smith's avatar
Ethan Smith committed
112
    def __post_init__(self) -> None:
Baber's avatar
Baber committed
113
114
115
116
117
118
        if self.output_type == "generate_until":
            if self.generation_kwargs is not None:
                if "temperature" in self.generation_kwargs:
                    self.generation_kwargs["temperature"] = float(
                        self.generation_kwargs["temperature"]
                    )
Lintang Sutawika's avatar
Lintang Sutawika committed
119

Baber's avatar
Baber committed
120
121
122
123
124
125
                if "until" not in self.generation_kwargs:
                    eval_logger.warning(
                        f"{self.task}: No `until` specified in `generation_kwargs`! Defaulting to the fewshot_delimiter={repr(self.fewshot_delimiter)}"
                    )
                    self.generation_kwargs["until"] = [self.fewshot_delimiter]
            else:
Lintang Sutawika's avatar
Lintang Sutawika committed
126
127
                # ensure that we greedily generate in absence of explicit arguments otherwise
                self.generation_kwargs = {
128
129
130
131
132
                    "until": (
                        None
                        if self.fewshot_delimiter is None
                        else [self.fewshot_delimiter]
                    ),
Lintang Sutawika's avatar
Lintang Sutawika committed
133
                    "do_sample": False,
Baber Abbasi's avatar
Baber Abbasi committed
134
                    "temperature": 0,
Lintang Sutawika's avatar
Lintang Sutawika committed
135
                }
Baber Abbasi's avatar
Baber Abbasi committed
136
137
138
                eval_logger.warning(
                    f"{self.task}: No `generation_kwargs` specified in task config, defaulting to {self.generation_kwargs}"
                )
Baber's avatar
Baber committed
139
140
141
142
143
        else:
            if self.generation_kwargs is not None:
                eval_logger.warning(
                    f"[{self.task}] passed `generation_kwargs`, but not using `output_type: generate_until`!"
                )
144

145
146
147
    def __getitem__(self, item):
        return getattr(self, item)

148
149
150
    def __setitem__(self, item, value):
        return setattr(self, item, value)

151
    def to_dict(self, keep_callable: bool = False) -> dict:
152
153
        """dumps the current config as a dictionary object, as a printable format.
        null fields will not be printed.
haileyschoelkopf's avatar
haileyschoelkopf committed
154
        Used for dumping results alongside full task configuration
155

haileyschoelkopf's avatar
haileyschoelkopf committed
156
157
158
159
160
161
162
163
164
165
        :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)
166
167
168
169
170
171
172
173
174
175
            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
176
        return cfg_dict
177

178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
    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)

194
195
196
197
198
199
200
201
202
203
204

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)
    """

205
    VERSION: Optional[Union[int, str]] = None
206

207
208
    # The name of the `Task` benchmark as denoted in the HuggingFace datasets Hub
    # or a path to a custom `datasets` loading script.
209
    DATASET_PATH: Optional[str] = None
210
211

    # The name of a subset within `DATASET_PATH`.
212
    DATASET_NAME: Optional[str] = None
213

214
    OUTPUT_TYPE: Optional[OutputType] = None
lintangsutawika's avatar
lintangsutawika committed
215

216
217
    def __init__(
        self,
218
219
220
221
        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
222
    ) -> None:
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
        """
        :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)
245
246
247
        self._training_docs: Optional[list] = None
        self._fewshot_docs: Optional[list] = None
        self._instances: Optional[List[Instance]] = None
248

249
        self._config: TaskConfig = TaskConfig({**config}) if config else TaskConfig()
250

lintangsutawika's avatar
lintangsutawika committed
251
        self._filters = [build_filter_ensemble("none", [["take_first", None]])]
252
253
254
        self.fewshot_rnd: Optional[random.Random] = (
            None  # purposely induce errors in case of improper usage
        )
255

256
257
258
259
260
261
    def download(
        self,
        data_dir: Optional[str] = None,
        cache_dir: Optional[str] = None,
        download_mode=None,
    ) -> None:
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
        """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.
        """
286
287
288
289
290
291
292
        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,
        )
293

294
    @property
295
    def config(self) -> TaskConfig:
296
297
298
        """Returns the TaskConfig associated with this class."""
        return self._config

299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
    @abc.abstractmethod
    def has_training_docs(self):
        """Whether the task has a training set"""
        pass

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

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

314
    def training_docs(self) -> Iterable:
315
316
317
318
319
320
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

321
    def validation_docs(self) -> Iterable:
322
323
324
325
326
327
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

328
    def test_docs(self) -> Iterable:
329
330
331
332
333
334
        """
        :return: Iterable[obj]
            A iterable of any object, that doc_to_text can handle
        """
        return []

335
    def fewshot_docs(self) -> Iterable:
336
337
338
339
340
341
342
343
344
        """
        :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
345
346
347
348
349
            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."
                )
350
351
            return self.test_docs()

352
    def _process_doc(self, doc: dict) -> dict:
353
354
355
356
357
358
359
360
361
        """
        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
362

363
    @property
364
    def instances(self) -> List[Instance]:
365
366
367
368
369
370
371
372
373
374
375
        """After calling `task.build_all_requests()`, tasks
        maintain a list of the dataset instances which will be evaluated.
        """
        return self._instances

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

        return rnd.sample(self._training_docs, k)

376
377
    def doc_to_decontamination_query(self, doc):
        raise NotImplementedError(
378
379
380
381
382
383
384
385
386
387
388
            "Override doc_to_decontamination_query with document specific decontamination query."
        )

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

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

389
390
391
392
    # not an abstractmethod because not every language-only task has to implement this
    def doc_to_image(self, doc):
        raise NotImplementedError

393
394
395
    def doc_to_audio(self, doc):
        raise NotImplementedError

Baber Abbasi's avatar
Baber Abbasi committed
396
397
398
    def doc_to_prefix(self, doc):
        return ""

399
400
    def build_all_requests(
        self,
401
        *,
402
        limit: Union[int, None] = None,
403
        samples: Optional[List[int]] = None,
404
405
406
407
408
409
410
411
412
        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 = "",
413
    ) -> None:
414
        """Build a set of Instances for a task, and store them in task.instances"""
415
416
417
418

        # used with caching
        og_limit = limit

419
        cache_key = f"requests-{self._config.task}-{self.config.num_fewshot}shot-rank{rank}-world_size{world_size}"
KonradSzafer's avatar
KonradSzafer committed
420
421
422
423
424
425
426
        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 ""
        )
427
        cache_key += f"-tokenizer{tokenizer_name}"
428

Baber Abbasi's avatar
Baber Abbasi committed
429
        cached_instances = load_from_cache(file_name=cache_key, cache=cache_requests)
430
431
432
433
434
435
436
437
438
439
440
441
442

        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
443
        eval_logger.info(f"Building contexts for {self.config.task} on rank {rank}...")
444

445
        instances = []
446
447
448
449
450
451
452
453
454
455

        # 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(
456
457
458
            self.doc_iterator(
                rank=rank, limit=limit, samples=samples, world_size=world_size
            )
459
460
461
462
463
464
465
        )

        num_docs = len(doc_id_docs)

        for doc_id, doc in tqdm(
            doc_id_docs,
            total=num_docs,
lintangsutawika's avatar
lintangsutawika committed
466
        ):
467
            # sample fewshot context #TODO: need to offset doc_id by rank now!
468
            fewshot_ctx = self.fewshot_context(
469
                doc,
470
471
472
473
474
475
476
                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
477
                gen_prefix=self.doc_to_prefix(doc),
478
            )
479

480
            # TODO: we should override self.config.repeats if doing greedy gen so users don't waste time+compute
lintangsutawika's avatar
lintangsutawika committed
481
482
483
            inst = self.construct_requests(
                doc=doc,
                ctx=fewshot_ctx,
484
                metadata=(self.config["task"], doc_id, self.config.repeats),
485
                apply_chat_template=apply_chat_template,
486
                chat_template=chat_template,
lintangsutawika's avatar
lintangsutawika committed
487
            )
488
489
490
491

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

492
493
494
495
496
497
498
499
500
501
502
503
504
            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
505

506
507
        if len(self._instances) == 0:
            raise ValueError("task.build_requests() did not find any docs!")
508

509
510
511
        if cache_requests and (not cached_instances or rewrite_requests_cache):
            save_to_cache(file_name=cache_key, obj=instances)

512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
    @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
528
            The number of times each instance in a dataset is inferred on. Defaults to 1,
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
            can be increased for techniques like majority voting.
        """
        pass

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

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

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

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

564
565
566
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

haileyschoelkopf's avatar
haileyschoelkopf committed
567
568
569
570
571
572
573
574
575
576
    @classmethod
    def count_bytes(cls, doc):
        """Used for byte-level perplexity metrics in rolling loglikelihood"""
        return len(doc.encode("utf-8"))

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

577
    @utils.positional_deprecated
Baber Abbasi's avatar
Baber Abbasi committed
578
    def fewshot_context(self, doc, num_fewshot, rnd=None, description=None, **kwargs):
579
580
581
582
583
584
585
        """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
586
587
588
589
590
        :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.
591
592
593
        :returns: str
            The fewshot context.
        """
594
        if rnd is None:
595
596
597
598
599
600
            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
601

602
        description = description if description else ""
603
604

        if num_fewshot == 0:
lintangsutawika's avatar
lintangsutawika committed
605
            labeled_examples = ""
606
        else:
lintangsutawika's avatar
lintangsutawika committed
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
            # 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
631
            )
632
633

        example = self.doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
634
        return description + labeled_examples + example
635

636
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
637
        """Iterates over FilterEnsembles and applies them to instances"""
lintangsutawika's avatar
lintangsutawika committed
638
639
        if hasattr(self, "_filters"):
            for f in self._filters:
640
                f.apply(self._instances)
lintangsutawika's avatar
lintangsutawika committed
641
642
643
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances
644

baberabb's avatar
baberabb committed
645
    def dump_config(self) -> dict:
Baber Abbasi's avatar
Baber Abbasi committed
646
        """Returns the config as a dictionary."""
647
        # TODO: this should only return the overrides applied to a non-YAML task's configuration.
648
        # (num_fewshot)
649
        return self.config.to_dict()
650

Baber Abbasi's avatar
Baber Abbasi committed
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
    def set_config(self, key: str, value: Any, update: bool = False) -> None:
        """Set or update the configuration for a given key."""
        if key is None:
            raise ValueError("Key must be provided.")

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

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

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

691
692
693
694
695
    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

696
697
698
699
700
701
702
    @property
    def eval_docs(self) -> Union[datasets.Dataset, List[dict]]:
        if self.has_test_docs():
            return self.test_docs()
        elif self.has_validation_docs():
            return self.validation_docs()
        else:
703
704
705
            raise ValueError(
                f"Task dataset (path={self.DATASET_PATH}, name={self.DATASET_NAME}) must have valid or test docs!"
            )
706
707

    def doc_iterator(
708
709
710
711
712
713
        self,
        *,
        rank: int = 0,
        limit: Union[int, None] = None,
        world_size: int = 1,
        samples: Optional[List[int]] = None,
714
    ) -> Iterator[Tuple[int, Any]]:
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
        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),
            )
737
738
        return doc_iterator

739
740

class ConfigurableTask(Task):
741
    VERSION = "Yaml"
742
    OUTPUT_TYPE = None
743
    CONFIG = None
744
745

    def __init__(
746
747
748
749
750
        self,
        data_dir=None,
        cache_dir=None,
        download_mode=None,
        config: Optional[dict] = None,
Ethan Smith's avatar
Ethan Smith committed
751
    ) -> None:  # TODO no super() call here
752
        # Get pre-configured attributes
753
        self._config = self.CONFIG
754

755
        # Use new configurations if there was no preconfiguration
756
        if self.config is None:
757
            self._config = TaskConfig(**config)
758
759
        # Overwrite configs
        else:
lintangsutawika's avatar
lintangsutawika committed
760
            if config is not None:
761
                self._config.__dict__.update(config)
762

763
        if self.config is None:
lintangsutawika's avatar
lintangsutawika committed
764
765
766
            raise ValueError(
                "Must pass a config to ConfigurableTask, either in cls.CONFIG or `config` kwarg"
            )
767

768
769
770
771
        if isinstance(self.config.metadata, dict):
            if "version" in self.config.metadata:
                self.VERSION = self.config.metadata["version"]

772
        if self.config.output_type is not None:
773
774
775
776
            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)}'"
                )
777
            self.OUTPUT_TYPE = self.config.output_type
778

779
780
781
782
        if self.config.doc_to_image is not None:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

783
784
785
786
        if self.config.doc_to_audio:
            # mark the task as requiring multimodality.
            self.MULTIMODAL = True

Hojin Lee's avatar
Hojin Lee committed
787
788
789
        if self.config.unsafe_code is not False:
            self.UNSAFE_CODE = True

790
791
        if self.config.dataset_path is not None:
            self.DATASET_PATH = self.config.dataset_path
792

793
794
        if self.config.dataset_name is not None:
            self.DATASET_NAME = self.config.dataset_name
795

Baber's avatar
Baber committed
796
797
        self.metric_results = []

798
799
800
801
        self._metric_fn_list = {}
        self._metric_fn_kwargs = {}
        self._aggregation_list = {}
        self._higher_is_better = {}
802

803
        if self.config.metric_list is None:
804
            # TODO: handle this in TaskConfig.__post_init__ ?
805
806
            _metric_list = DEFAULT_METRIC_REGISTRY[self.config.output_type]

807
            for metric_name in _metric_list:
haileyschoelkopf's avatar
haileyschoelkopf committed
808
                self._metric_fn_list[metric_name] = get_metric(metric_name)
lintangsutawika's avatar
lintangsutawika committed
809
                self._metric_fn_kwargs[metric_name] = {}
810
811
812
                self._aggregation_list[metric_name] = get_metric_aggregation(
                    metric_name
                )
haileyschoelkopf's avatar
haileyschoelkopf committed
813
                self._higher_is_better[metric_name] = is_higher_better(metric_name)
814
        else:
815
            for metric_config in self.config.metric_list:
816
817
818
819
                if "metric" not in metric_config:
                    raise ValueError(
                        "'metric' key not provided for an entry in 'metric_list', must be specified!"
                    )
820
821
822
823
                metric_name = metric_config["metric"]
                kwargs = {
                    key: metric_config[key]
                    for key in metric_config
Chris's avatar
Chris committed
824
825
                    if key
                    not in ["metric", "aggregation", "higher_is_better", "hf_evaluate"]
826
                }
Chris's avatar
Chris committed
827
828
829
830
                hf_evaluate_metric = (
                    "hf_evaluate" in metric_config
                    and metric_config["hf_evaluate"] is True
                )
831

832
                if self.config.process_results is not None:
833
834
                    self._metric_fn_list[metric_name] = None
                    self._metric_fn_kwargs[metric_name] = {}
835
836
837
838
839
840
                elif callable(metric_name):
                    metric_fn = metric_name.__call__
                    metric_name = metric_name.__name__
                    self._metric_fn_list[metric_name] = metric_fn
                    self._metric_fn_kwargs[metric_name] = kwargs
                else:
Chris's avatar
Chris committed
841
842
843
                    self._metric_fn_list[metric_name] = get_metric(
                        metric_name, hf_evaluate_metric
                    )
844
                    self._metric_fn_kwargs[metric_name] = kwargs
lintangsutawika's avatar
lintangsutawika committed
845

846
                if "aggregation" in metric_config:
847
                    agg_name = metric_config["aggregation"]
848
                    if isinstance(agg_name, str):
haileyschoelkopf's avatar
haileyschoelkopf committed
849
                        self._aggregation_list[metric_name] = get_aggregation(agg_name)
850
                    elif callable(agg_name):  # noqa: E721
851
852
853
                        self._aggregation_list[metric_name] = metric_config[
                            "aggregation"
                        ]
854
                else:
855
                    INV_AGG_REGISTRY = {v: k for k, v in AGGREGATION_REGISTRY.items()}
lintangsutawika's avatar
lintangsutawika committed
856
                    metric_agg = get_metric_aggregation(metric_name)
857
                    eval_logger.warning(
858
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but aggregation is not. "
859
860
                        f"using default "
                        f"aggregation={INV_AGG_REGISTRY[metric_agg]}"
861
                    )
862
                    self._aggregation_list[metric_name] = metric_agg
lintangsutawika's avatar
lintangsutawika committed
863

864
865
866
867
868
869
                if "higher_is_better" in metric_config:
                    self._higher_is_better[metric_name] = metric_config[
                        "higher_is_better"
                    ]
                else:
                    eval_logger.warning(
870
                        f"[Task: {self.config.task}] metric {metric_name} is defined, but higher_is_better is not. "
871
                        f"using default "
haileyschoelkopf's avatar
haileyschoelkopf committed
872
                        f"higher_is_better={is_higher_better(metric_name)}"
873
                    )
haileyschoelkopf's avatar
haileyschoelkopf committed
874
                    self._higher_is_better[metric_name] = is_higher_better(metric_name)
875

876
        self.download(self.config.dataset_kwargs)
877
878
879
        self._training_docs = None
        self._fewshot_docs = None

880
        if self.config.filter_list is not None:
lintangsutawika's avatar
lintangsutawika committed
881
            self._filters = []
882
            for filter_config in self.config.filter_list:
883
884
885
886
887
888
889
890
891
                filter_name = filter_config["name"]
                filter_functions = filter_config["filter"]
                components = []
                for function in filter_functions:
                    kwargs = {
                        key: function[key] for key in function if key != "function"
                    }
                    components.append([function["function"], kwargs])
                filter_pipeline = build_filter_ensemble(filter_name, components)
lintangsutawika's avatar
lintangsutawika committed
892
                self._filters.append(filter_pipeline)
lintangsutawika's avatar
lintangsutawika committed
893
        else:
Baber Abbasi's avatar
Baber Abbasi committed
894
895
896
897
            # 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."
            )
898
            self._filters = [build_filter_ensemble("none", [["take_first", None]])]
899

900
901
        if self.config.use_prompt is not None:
            eval_logger.info(f"loading prompt {self.config.use_prompt}")
902
            self.prompt = get_prompt(
903
                self.config.use_prompt, self.DATASET_PATH, self.DATASET_NAME
lintangsutawika's avatar
lintangsutawika committed
904
            )
905
906
907
        else:
            self.prompt = None

lintangsutawika's avatar
lintangsutawika committed
908
        if self.fewshot_docs() is not None:
909
910
911
912
            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
913
914
915
                self.config.fewshot_config.get("sampler", "default")
                if self.config.fewshot_config
                else "default"
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
            )
            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)}"
                )
932

933
        self.task_docs = self.eval_docs
934

935
        # Test One Doc
936
        self.features = list(self.task_docs.features.keys())
937
938
        self.multiple_input = 0
        self.multiple_target = 0
939
        test_doc = self.task_docs[0]
940
        test_text = self.doc_to_text(test_doc)
941
        test_target = self.doc_to_target(test_doc)
lintangsutawika's avatar
lintangsutawika committed
942

943
        if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
944
            test_choice = self.doc_to_choice(test_doc)
945
            if not isinstance(test_choice, list):
lintangsutawika's avatar
lintangsutawika committed
946
                eval_logger.error("doc_to_choice must return list")
947
948
            else:
                num_choice = len(test_choice)
949

950
            if isinstance(test_text, int):
Baber Abbasi's avatar
Baber Abbasi committed
951
952
953
                eval_logger.debug(
                    "doc_to_text returned an int. Assuming multiple inputs."
                )
954
                self.multiple_input = num_choice
955
956
        else:
            test_choice = None
957

958
        if isinstance(test_target, list):
Baber Abbasi's avatar
Baber Abbasi committed
959
960
961
            eval_logger.debug(
                "doc_to_target returned a list. Assuming multiple targets."
            )
962
            self.multiple_target = len(test_target)
963
        else:
964
            if (isinstance(test_target, int)) and (test_choice is not None):
lintangsutawika's avatar
lintangsutawika committed
965
                test_target = test_choice[test_target]
966
            else:
lintangsutawika's avatar
lintangsutawika committed
967
                test_target = str(test_target)
968

969
970
971
        if test_choice is not None:
            check_choices = test_choice
        else:
lintangsutawika's avatar
lintangsutawika committed
972
            check_choices = [test_target]
973
974
975
976
        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 = (
977
978
                    True
                    if self.config.target_delimiter.rstrip()
979
                    != self.config.target_delimiter
980
                    else False
981
                )
982

983
                if delimiter_has_whitespace and choice_has_whitespace:
984
985
                    eval_logger.debug(
                        f'Both target_delimiter "{self.config.target_delimiter}" and target choice: "{choice}" have whitespace'
986
987
                    )
                elif (not delimiter_has_whitespace) and (not choice_has_whitespace):
988
                    eval_logger.debug(
989
                        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'
990
991
                    )

Baber Abbasi's avatar
Baber Abbasi committed
992
993
994
995
996
997
998
999
1000
1001
1002
1003
1004
1005
1006
1007
1008
    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 {},
            )
1009

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

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

baberabb's avatar
baberabb committed
1022
    def has_test_docs(self) -> bool:
1023
        if self.config.test_split is not None:
1024
1025
1026
1027
            return True
        else:
            return False

baberabb's avatar
baberabb committed
1028
    def training_docs(self) -> datasets.Dataset:
1029
        if self.has_training_docs():
1030
1031
1032
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.training_split]
1033
                )
1034
            return self.dataset[self.config.training_split]
1035

baberabb's avatar
baberabb committed
1036
    def validation_docs(self) -> datasets.Dataset:
1037
        if self.has_validation_docs():
1038
1039
1040
            if self.config.process_docs is not None:
                return self.config.process_docs(
                    self.dataset[self.config.validation_split]
1041
                )
1042
            return self.dataset[self.config.validation_split]
1043

baberabb's avatar
baberabb committed
1044
    def test_docs(self) -> datasets.Dataset:
1045
        if self.has_test_docs():
1046
1047
1048
            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]
1049

1050
    def fewshot_docs(self):
1051
        if self.config.fewshot_split is not None:
1052
1053
            if self.config.process_docs is not None:
                return self.config.process_docs(self.dataset[self.config.fewshot_split])
1054
            return self.dataset[self.config.fewshot_split]
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
        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."
                )
1067
        else:
1068
            if (self.config.num_fewshot is not None) and (self.config.num_fewshot > 0):
1069
                eval_logger.warning(
Lintang Sutawika's avatar
Lintang Sutawika committed
1070
                    f"[Task: {self.config.task}] "
1071
1072
1073
1074
                    "num_fewshot > 0 but fewshot_split is None. "
                    "using preconfigured rule."
                )
            return super().fewshot_docs()
1075

KonradSzafer's avatar
KonradSzafer committed
1076
1077
1078
1079
1080
    @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
1081
        gen_prefix: Optional[str] = None,
KonradSzafer's avatar
KonradSzafer committed
1082
1083
1084
1085
1086
1087
1088
1089
    ) -> 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":
1090
                labeled_examples.append({"role": "user", "content": question})
KonradSzafer's avatar
KonradSzafer committed
1091
1092
            # if last message is user, append to it to avoid two user messages in a row
            else:
1093
                labeled_examples[-1]["content"] += question
KonradSzafer's avatar
KonradSzafer committed
1094
1095
        else:
            # if fewshot_as_multiturn is True, append as next user entry (last is always assistant)
1096
            labeled_examples.append({"role": "user", "content": question})
Baber Abbasi's avatar
Baber Abbasi committed
1097
1098
        if gen_prefix:
            labeled_examples.append({"role": "assistant", "content": gen_prefix})
KonradSzafer's avatar
KonradSzafer committed
1099

lintangsutawika's avatar
lintangsutawika committed
1100
    @utils.positional_deprecated
KonradSzafer's avatar
KonradSzafer committed
1101
1102
    def fewshot_context(
        self,
Baber Abbasi's avatar
Baber Abbasi committed
1103
        doc: dict,
KonradSzafer's avatar
KonradSzafer committed
1104
1105
1106
1107
        num_fewshot: int,
        system_instruction: Optional[str] = None,
        apply_chat_template: bool = False,
        fewshot_as_multiturn: bool = False,
1108
        chat_template: Optional[Callable] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1109
        gen_prefix: Optional[str] = None,
Baber Abbasi's avatar
Baber Abbasi committed
1110
    ) -> Union[str, List[str]]:
lintangsutawika's avatar
lintangsutawika committed
1111
1112
1113
1114
1115
1116
1117
        """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
1118
1119
1120
1121
1122
1123
        :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.
1124
1125
        :param chat_template:
            callable (from lm.apply_chat_template) that takes in a list[Dict] chat transcript and renders it into a string.
1126
1127
        :param gen_prefix:
            String to append after the <|assistant|> token.
lintangsutawika's avatar
lintangsutawika committed
1128
1129
1130
        :returns: str
            The fewshot context.
        """
KonradSzafer's avatar
KonradSzafer committed
1131
1132
1133
1134
1135
1136
        if apply_chat_template:
            labeled_examples = []
        else:
            labeled_examples = ""

        # get task description
1137
1138
        if description := self.config.description:
            description = utils.apply_template(self.config.description, doc)
lintangsutawika's avatar
lintangsutawika committed
1139

KonradSzafer's avatar
KonradSzafer committed
1140
1141
1142
1143
1144
1145
1146
1147
1148
        # 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
1149
        else:
KonradSzafer's avatar
KonradSzafer committed
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
            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
1163
1164
1165
                        doc,
                        num_fewshot,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1166
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1167
1168
1169
                    )
                )
            else:
Baber Abbasi's avatar
Baber Abbasi committed
1170
                labeled_examples += self.sampler.get_context(
Baber Abbasi's avatar
Baber Abbasi committed
1171
                    doc, num_fewshot, gen_prefix=gen_prefix
Baber Abbasi's avatar
Baber Abbasi committed
1172
                )
lintangsutawika's avatar
lintangsutawika committed
1173
1174

        example = self.doc_to_text(doc)
KonradSzafer's avatar
KonradSzafer committed
1175
1176
        if apply_chat_template:
            if self.multiple_input:
Baber Abbasi's avatar
Baber Abbasi committed
1177
                # TODO: append prefill?
1178
1179
                if not labeled_examples:
                    return ""
1180
                return chat_template(labeled_examples)
KonradSzafer's avatar
KonradSzafer committed
1181
1182
            if isinstance(example, str):
                self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1183
1184
1185
                    labeled_examples,
                    example,
                    fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1186
                    gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1187
1188
1189
1190
1191
1192
1193
                )
            # 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
1194
1195
1196
1197
                    self.append_target_question(
                        chat,
                        ex,
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1198
                        gen_prefix=gen_prefix,
Baber Abbasi's avatar
Baber Abbasi committed
1199
1200
1201
1202
1203
                    )
                    # TODO: append prefill?
                    labeled_examples_list.append(
                        chat_template(
                            chat,
Baber Abbasi's avatar
Baber Abbasi committed
1204
                            add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1205
1206
                        )
                    )
KonradSzafer's avatar
KonradSzafer committed
1207
1208
1209
1210
1211
1212
                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
1213
1214
1215
                        labeled_examples,
                        choices[example],
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1216
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1217
1218
1219
                    )
                else:
                    self.append_target_question(
Baber Abbasi's avatar
Baber Abbasi committed
1220
1221
1222
                        labeled_examples,
                        str(example),
                        fewshot_as_multiturn,
Baber Abbasi's avatar
Baber Abbasi committed
1223
                        gen_prefix=gen_prefix,
KonradSzafer's avatar
KonradSzafer committed
1224
1225
                    )
                # return lm.apply_chat_template(labeled_examples)
Baber Abbasi's avatar
Baber Abbasi committed
1226
1227
            return chat_template(
                labeled_examples,
Baber Abbasi's avatar
Baber Abbasi committed
1228
                add_generation_prompt=False if gen_prefix else True,
Baber Abbasi's avatar
Baber Abbasi committed
1229
            )
1230
        else:
Baber Abbasi's avatar
Baber Abbasi committed
1231
            prefix = (
Baber Abbasi's avatar
Baber Abbasi committed
1232
1233
                self.config.target_delimiter + gen_prefix
                if gen_prefix is not None
Baber Abbasi's avatar
Baber Abbasi committed
1234
1235
                else ""
            )
KonradSzafer's avatar
KonradSzafer committed
1236
1237
            if self.multiple_input:
                return labeled_examples
1238
            if isinstance(example, str):
Baber Abbasi's avatar
Baber Abbasi committed
1239
                return labeled_examples + example + prefix
1240
            elif isinstance(example, list):
Baber Abbasi's avatar
Baber Abbasi committed
1241
                return [labeled_examples + ex + prefix for ex in example]
1242
1243
1244
            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
1245
                    return labeled_examples + choices[example] + prefix
1246
                else:
Baber Abbasi's avatar
Baber Abbasi committed
1247
                    return labeled_examples + str(example) + prefix
lintangsutawika's avatar
lintangsutawika committed
1248

Baber Abbasi's avatar
Baber Abbasi committed
1249
    def apply_filters(self) -> Optional[List[Instance]]:
Baber Abbasi's avatar
Baber Abbasi committed
1250
        """Iterates over FilterEnsembles and applies them to instances"""
1251
1252
        if hasattr(self, "_filters"):
            for f in self._filters:
1253
                f.apply(self._instances)
1254
1255
1256
1257
        else:
            eval_logger.warning("No filter defined, passing through instances")
            return self._instances

1258
    def should_decontaminate(self):
1259
        return self.config.should_decontaminate
1260

Baber Abbasi's avatar
Baber Abbasi committed
1261
    def doc_to_decontamination_query(self, doc: dict):
1262
        if self.config.should_decontaminate:
1263
1264
            if self.config.doc_to_decontamination_query is None:
                return self.doc_to_text(doc)
1265
            else:
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
                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
                        )
                    )
1277

1278
    def _process_doc(self, doc: dict) -> dict:
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
        """
        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
1289
    def doc_to_text(self, doc, doc_to_text=None):
1290
1291
        if self.prompt is not None:
            doc_to_text = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1292
1293
        elif doc_to_text is not None:
            doc_to_text = doc_to_text
1294
        else:
1295
            doc_to_text = self.config.doc_to_text
lintangsutawika's avatar
lintangsutawika committed
1296

1297
        if isinstance(doc_to_text, int):
1298
            return doc_to_text
1299
        elif isinstance(doc_to_text, str):
1300
            if doc_to_text in self.features:
1301
                # if self.config.doc_to_choice is not None:
1302
1303
                #     return self.doc_to_choice(doc)[doc[doc_to_text]]
                # else:
1304
1305
                return doc[doc_to_text]
            else:
lintangsutawika's avatar
lintangsutawika committed
1306
                text_string = utils.apply_template(doc_to_text, doc)
lintangsutawika's avatar
lintangsutawika committed
1307
                if text_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1308
1309
1310
                    return ast.literal_eval(text_string)
                else:
                    return text_string
1311
        elif callable(doc_to_text):
1312
            return doc_to_text(doc)
lintangsutawika's avatar
lintangsutawika committed
1313
        # Used when applying a Promptsource template
1314
        elif hasattr(doc_to_text, "apply"):
1315
1316
1317
1318
1319
            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")
1320
                return self.config.fewshot_delimiter
1321
        else:
1322
            print(type(doc_to_text))
1323
            raise TypeError
1324

Yu Shi Jie's avatar
Yu Shi Jie committed
1325
    def doc_to_target(self, doc: Mapping, doc_to_target=None) -> Union[int, str, list]:
1326
1327
        if self.prompt is not None:
            doc_to_target = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1328
1329
        elif doc_to_target is not None:
            doc_to_target = doc_to_target
1330
        else:
1331
            doc_to_target = self.config.doc_to_target
1332

1333
        if isinstance(doc_to_target, int):
1334
            return doc_to_target
1335
        elif isinstance(doc_to_target, str):
1336
            if doc_to_target in self.features:
1337
                # if self.config.doc_to_choice is not None:
1338
1339
1340
                #     return self.doc_to_choice(doc)[doc[doc_to_target]]
                # else:
                return doc[doc_to_target]
1341
            else:
lintangsutawika's avatar
lintangsutawika committed
1342
                target_string = utils.apply_template(doc_to_target, doc)
lintangsutawika's avatar
lintangsutawika committed
1343
                if target_string.isdigit() and self._config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1344
                    return ast.literal_eval(target_string)
lintangsutawika's avatar
lintangsutawika committed
1345
1346
1347
1348
1349
                elif (
                    len(target_string) >= 2
                    and (target_string[0] == "[")
                    and (target_string[-1] == "]")
                ):
baberabb's avatar
baberabb committed
1350
1351
1352
1353
                    try:
                        return ast.literal_eval(target_string)
                    except (SyntaxError, ValueError):
                        return target_string
lintangsutawika's avatar
lintangsutawika committed
1354
1355
                else:
                    return target_string
1356
        elif isinstance(doc_to_target, list):
1357
            return doc_to_target
1358
        elif callable(doc_to_target):
1359
            return doc_to_target(doc)
lintangsutawika's avatar
lintangsutawika committed
1360
        # Used when applying a Promptsource template
1361
        elif hasattr(doc_to_target, "apply"):
1362
            applied_prompt = doc_to_target.apply(doc)
1363
1364
1365
1366
            if len(applied_prompt) == 2:
                return applied_prompt[1]
            else:
                eval_logger.warning("Applied prompt returns empty string")
1367
                return self.config.fewshot_delimiter
1368
1369
        else:
            raise TypeError
1370

Yu Shi Jie's avatar
Yu Shi Jie committed
1371
    def doc_to_choice(self, doc: Any, doc_to_choice=None) -> List[str]:
1372
1373
        if self.prompt is not None:
            doc_to_choice = self.prompt
Yu Shi Jie's avatar
Yu Shi Jie committed
1374
1375
        elif doc_to_choice is not None:
            doc_to_choice = doc_to_choice
1376
        elif self.config.doc_to_choice is None:
1377
1378
            eval_logger.error("doc_to_choice was called but not set in config")
        else:
1379
            doc_to_choice = self.config.doc_to_choice
1380

1381
        if isinstance(doc_to_choice, str):
1382
1383
1384
1385
            if doc_to_choice in self.features:
                return doc[doc_to_choice]
            else:
                return ast.literal_eval(utils.apply_template(doc_to_choice, doc))
1386
        elif isinstance(doc_to_choice, list):
1387
            return doc_to_choice
1388
        elif isinstance(doc_to_choice, dict):
1389
1390
1391
1392
1393
1394
1395
            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
1396

1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
    def doc_to_image(self, doc: Any, doc_to_image=None) -> Union[int, str, list]:
        if doc_to_image is not None:
            doc_to_image = doc_to_image
        elif self.config.doc_to_image is not None:
            doc_to_image = self.config.doc_to_image
        else:
            return None

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

1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
    def doc_to_audio(self, doc: Any, doc_to_audio=None) -> Union[int, str, list]:
        if doc_to_audio is not None:
            doc_to_audio = doc_to_audio
        elif self.config.doc_to_audio is not None:
            doc_to_audio = self.config.doc_to_audio
        else:
            return None

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

Baber Abbasi's avatar
Baber Abbasi committed
1443
1444
1445
1446
1447
1448
1449
1450
    def doc_to_prefix(self, doc):
        if (gen_prefix := self.config.gen_prefix) is not None:
            if gen_prefix in self.features:
                return doc[gen_prefix]
            else:
                return utils.apply_template(gen_prefix, doc)
        return None

baberabb's avatar
baberabb committed
1451
1452
1453
    def construct_requests(
        self, doc: dict, ctx: str, **kwargs
    ) -> Union[List[Instance], Instance]:
1454
        apply_chat_template = kwargs.pop("apply_chat_template", False)
1455
        chat_template: Callable | None = kwargs.pop("chat_template", None)
1456

1457
1458
        aux_arguments = None

1459
        if self.OUTPUT_TYPE == "loglikelihood":
lintangsutawika's avatar
lintangsutawika committed
1460
            arguments = (ctx, self.doc_to_target(doc))
1461
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
lintangsutawika's avatar
lintangsutawika committed
1462
            arguments = (self.doc_to_target(doc),)
1463
        elif self.OUTPUT_TYPE == "multiple_choice":
1464
            choices = self.doc_to_choice(doc)
1465
            target_delimiter = self.config.target_delimiter
1466
1467
            if apply_chat_template:
                target_delimiter = ""
1468
1469
            if self.multiple_input:
                # If there are multiple inputs, choices are placed in the ctx
1470
                # apply chat_template to choices if apply_chat_template
1471
                cont = self.doc_to_target(doc)
1472

1473
                arguments = [
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
                    (
                        ctx
                        + (
                            chat_template([{"role": "user", "content": choice}])
                            if apply_chat_template
                            else choice
                        ),
                        f"{target_delimiter}{cont}",
                    )
                    for choice in choices
1484
                ]
1485
            else:
1486
                # Otherwise they are placed in the continuation
1487
                arguments = [(ctx, f"{target_delimiter}{cont}") for cont in choices]
1488

1489
1490
1491
1492
1493
1494
1495
1496
            # TODO: we should raise a warning telling users this will at most ~2x runtime.
            if "acc_mutual_info" in self._metric_fn_list.keys():
                # if we are calculating multiple choice accuracy
                # using mutual information instead of raw loglikelihood as metric, need unconditional lls.

                # here mutual info refers to calculating
                # log(P(choice|ctx) / P(choice)) = log(P(choice|ctx)) - log(P(choice))
                # in other words normalizing by subtracting the unconditional logprob of each choice.
1497
1498
1499
1500
                # 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
                ]
1501
1502
1503
1504
1505
1506

                arguments.extend(aux_arguments)

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

Baber's avatar
Baber committed
1507
1508
1509
1510
1511
1512
1513
        else:
            raise ValueError(
                f"Unsupported OUTPUT_TYPE: '{self.OUTPUT_TYPE}'. "
                f"Expected one of: 'loglikelihood', 'loglikelihood_rolling', "
                f"'multiple_choice', 'generate_until'"
            )

1514
1515
1516
1517
1518
1519
1520
1521
1522
        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)},
            }

1523
1524
1525
1526
1527
1528
1529
1530
        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)},
            }

1531
1532
1533
1534
1535
1536
1537
        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":
1538
            request_list = [
1539
1540
                Instance(
                    request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1541
                    doc=doc,
Baber's avatar
Baber committed
1542
1543
                    arguments=arg,
                    # arguments=LoglikelihoodInput(context=arg[0], continuation=arg[1]),
1544
                    idx=i,
1545
1546
                    **kwargs,
                )
1547
                for i, arg in enumerate(arguments)
1548
            ]
1549
1550

            return request_list
lintangsutawika's avatar
lintangsutawika committed
1551

lintangsutawika's avatar
lintangsutawika committed
1552
        return Instance(
1553
1554
            request_type=self.OUTPUT_TYPE,
            doc=doc,
Baber's avatar
Baber committed
1555
1556
1557
            arguments=arguments,
            # if self.OUTPUT_TYPE in ["loglikelihood", "loglikelihood_rolling"]
            # else GenerateInput(*arguments),
1558
1559
            idx=0,
            **kwargs,
lintangsutawika's avatar
lintangsutawika committed
1560
        )
1561

Baber's avatar
Baber committed
1562
    def process_results(self, doc, results) -> dict:
1563
1564
        if callable(self.config.process_results):
            return self.config.process_results(doc, results)
lintangsutawika's avatar
lintangsutawika committed
1565

1566
        result_dict = {}
1567
        use_metric = list(self._metric_fn_list.keys())
1568
1569
1570
        if self.OUTPUT_TYPE == "loglikelihood":
            results = results[0]
            ll, is_greedy = results
1571
1572
1573
1574
            return {
                **({"perplexity": ll} if "perplexity" in use_metric else {}),
                **({"acc": int(is_greedy)} if "acc" in use_metric else {}),
            }
1575
        elif self.OUTPUT_TYPE == "loglikelihood_rolling":
haileyschoelkopf's avatar
haileyschoelkopf committed
1576
            (loglikelihood,) = results
1577
1578
            _words = self.count_words(self.doc_to_target(doc))
            _bytes = self.count_bytes(self.doc_to_target(doc))
haileyschoelkopf's avatar
haileyschoelkopf committed
1579
            return {
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
                **(
                    {"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
1595
            }
1596
        elif self.OUTPUT_TYPE == "multiple_choice":
1597
            lls, is_greedy = zip(*results)
lintangsutawika's avatar
lintangsutawika committed
1598

1599
            # retrieve choices in List[str] form, to compute choice lengths, etc.
1600
            choices = self.doc_to_choice(doc)
1601
1602
            completion_len = np.array([float(len(i)) for i in choices])

1603
1604
            if (
                2 * len(choices) == len(lls)
1605
                and "acc_mutual_info" in self._metric_fn_list.keys()
1606
1607
1608
            ):
                # then we are doing mutual info.
                # this stores the "dryrun" / unconditional answer loglikelihoods
1609
1610
                # as we extend the args list with unconditional ("", continuation) pairs
                lls_unconditional = lls[len(choices) :]
1611
1612
                if len(lls_unconditional) != len(choices):
                    raise ValueError
1613
                # and this stores our "regular" conditional loglikelihoods
1614
                lls = lls[: len(choices)]
1615

1616
1617
            pred = np.argmax(lls)
            pred_norm = np.argmax(lls / completion_len)
lintangsutawika's avatar
lintangsutawika committed
1618

1619
1620
            if self.multiple_input:
                gold = self.doc_to_text(doc)
1621
            else:
1622
                gold = self.doc_to_target(doc)
1623
1624

            gold_index_error = False
1625
            if isinstance(gold, list):
Lintang Sutawika's avatar
Lintang Sutawika committed
1626
1627
                gold = [i if i < len(choices) else -100 for i in gold]
                if -100 in gold:
1628
1629
                    gold_index_error = True
            else:
1630
                if isinstance(gold, int):
Lintang Sutawika's avatar
Lintang Sutawika committed
1631
                    gold = gold if gold < len(choices) else -100
1632
                elif isinstance(gold, str):
Lintang Sutawika's avatar
Lintang Sutawika committed
1633
                    gold = choices.index(gold) if gold in choices else -100
lintangsutawika's avatar
lintangsutawika committed
1634

Lintang Sutawika's avatar
Lintang Sutawika committed
1635
                if gold == -100:
1636
1637
1638
1639
                    gold_index_error = True

            if gold_index_error:
                eval_logger.warning(
lintangsutawika's avatar
lintangsutawika committed
1640
                    f"Label index was not in within range of available choices,"
1641
1642
                    f"Sample:\n\n{doc}\n\n"
                )
lintangsutawika's avatar
lintangsutawika committed
1643

1644
            if self.multiple_target:
lintangsutawika's avatar
lintangsutawika committed
1645
1646
                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
1647
                exact_match = int(any([is_greedy[i] if i != -100 else 0 for i in gold]))
lintangsutawika's avatar
lintangsutawika committed
1648
1649
1650
            else:
                acc = 1.0 if pred == gold else 0.0
                acc_norm = 1.0 if pred_norm == gold else 0.0
1651
                # 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
1652
                exact_match = int(is_greedy[gold]) if gold != -100 else 0
1653

Lintang Sutawika's avatar
Lintang Sutawika committed
1654
1655
1656
1657
            prob_norm = utils.softmax(lls)

            # TODO use keyword arguments to the metric?
            # gold, pred, norm stuff, the original lls,
1658
            result_dict = {
1659
                **({"acc": acc} if "acc" in use_metric else {}),
1660
1661
                **({"f1": (gold, pred)} if "f1" in use_metric else {}),
                **({"mcc": (gold, pred)} if "mcc" in use_metric else {}),
1662
                **({"acc_norm": acc_norm} if "acc_norm" in use_metric else {}),
1663
                **({"exact_match": exact_match} if "exact_match" in use_metric else {}),
Lintang Sutawika's avatar
Lintang Sutawika committed
1664
1665
1666
1667
1668
                **(
                    {"brier_score": (gold, prob_norm)}
                    if "brier_score" in use_metric
                    else {}
                ),
1669
1670
            }

1671
            if "acc_mutual_info" in use_metric:
lintangsutawika's avatar
lintangsutawika committed
1672
1673
1674
                lls_mutual_info = [
                    ll_c - ll_u for ll_c, ll_u in zip(lls, lls_unconditional)
                ]
1675
1676
1677
                acc_mutual_info = 1.0 if np.argmax(lls_mutual_info) == gold else 0.0
                result_dict["acc_mutual_info"] = acc_mutual_info

1678
        elif self.OUTPUT_TYPE == "generate_until":
1679
            gold = self.doc_to_target(doc)
Chris's avatar
Chris committed
1680
            result = results[0]
1681
            if self.config.doc_to_choice is not None:
lintangsutawika's avatar
lintangsutawika committed
1682
                # If you set doc_to_choice,
lintangsutawika's avatar
lintangsutawika committed
1683
                # it assumes that doc_to_target returns a number.
1684
1685
                choices = self.doc_to_choice(doc)
                gold = choices[gold]
1686
1687
            # we expect multiple_targets to be a list.
            elif self.multiple_target:
baberabb's avatar
baberabb committed
1688
                gold = list(gold)
Hojin Lee's avatar
Hojin Lee committed
1689
1690
1691
            # TODO: handle this better
            elif type(gold) is not type(result) and not (
                "bypass" in self._metric_fn_list.keys() or isinstance(result, list)
1692
            ):
Chris's avatar
Chris committed
1693
1694
                # cast gold to the same type as result
                gold = type(result)(gold)
1695

lintangsutawika's avatar
lintangsutawika committed
1696
            for metric in self._metric_fn_list.keys():
haileyschoelkopf's avatar
haileyschoelkopf committed
1697
1698
1699
1700
1701
                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
1702
1703
1704
1705
                    if not isinstance(gold, list):
                        # sometimes, a multiple_target dataset has exceptions where one doc has only one string answer
                        # print(gold)
                        gold = [gold]
1706
1707
1708
1709
1710
1711
1712
1713
                    if metric == "exact_match":
                        result = [result for _ in range(len(gold))]
                        scores = self._metric_fn_list[metric](
                            references=gold,
                            predictions=result,
                            **self._metric_fn_kwargs[metric],
                        )[metric]
                        result_score = 1.0 if scores > 0.0 else 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1714
                    else:
1715
1716
1717
1718
1719
1720
1721
1722
1723
1724
1725
1726
1727
1728
1729
1730
1731
1732
1733
1734
1735
                        for gold_option in gold:
                            try:
                                result_score = self._metric_fn_list[metric](
                                    references=[gold_option],
                                    predictions=[result],
                                    **self._metric_fn_kwargs[metric],
                                )
                            except (
                                TypeError
                            ):  # TODO: this is hacky and I don't want to do it
                                result_score = self._metric_fn_list[metric](
                                    [gold_option, result]
                                )
                            if isinstance(result_score, dict):
                                # TODO: this handles the case where HF evaluate returns a dict.
                                result_score = result_score[metric]
                            scores.append(result_score)
                        if any(scores):
                            result_score = 1.0
                        else:
                            result_score = 0.0
haileyschoelkopf's avatar
haileyschoelkopf committed
1736
                else:
1737
                    try:
1738
                        result_score = self._metric_fn_list[metric](
1739
1740
                            references=[gold],
                            predictions=[result],
1741
                            **self._metric_fn_kwargs[metric],
1742
                        )
1743
                    except TypeError:  # needed for now in order to use a different interface between our own metrics and HF Evaluate metrics
1744
                        result_score = self._metric_fn_list[metric]([gold, result])
1745
1746
1747
1748
1749
1750
1751
                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
1752
        else:
lintangsutawika's avatar
lintangsutawika committed
1753
1754
            raise ValueError(
                f"Passed invalid output_type '{self.OUTPUT_TYPE}' ! Please use one of ",
1755
                "'loglikelihood', 'loglikelihood_rolling', 'generate_until' or 'multiple_choice'",
1756
            )
1757
1758
1759

        return result_dict

Baber Abbasi's avatar
Baber Abbasi committed
1760
    def aggregation(self) -> dict:
1761
1762
        return self._aggregation_list

Baber Abbasi's avatar
Baber Abbasi committed
1763
    def higher_is_better(self) -> dict:
haileyschoelkopf's avatar
haileyschoelkopf committed
1764
        return self._higher_is_better
1765

Baber Abbasi's avatar
Baber Abbasi committed
1766
1767
1768
    def get_config(self, key: str) -> Any:
        return getattr(self._config, key, None)

Lintang Sutawika's avatar
Lintang Sutawika committed
1769
1770
1771
1772
    @property
    def task_name(self) -> Any:
        return getattr(self.config, "task", None)

1773
1774
1775
1776
1777
    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
1778
            f"num_samples={len(self.eval_docs)})"
1779
1780
        )

Baber's avatar
rename  
Baber committed
1781
    def compute_sample_metrics(
Baber's avatar
TODO!  
Baber committed
1782
        self,
Baber's avatar
Baber committed
1783
1784
1785
        requests: Optional[list[Instance]] = None,
        filter_keys: Optional[list[str]] = None,
        indices: Optional[list[int]] = None,
Baber's avatar
Baber committed
1786
        rank: int = 1,
Baber's avatar
Baber committed
1787
        limit: Optional[int] = None,
Baber's avatar
Baber committed
1788
1789
1790
1791
1792
        world_size: int = 1,
        log_samples: bool = False,
    ) -> tuple[
        Optional[dict[tuple[str, str], list[list[float]]]], Optional[list[dict]]
    ]:
1793
1794
1795
1796
1797
        """Calculate metrics for all datapoints in the task.

        Args:
            instances_by_doc_id (dict): Dictionary mapping doc_ids to lists of instances.
            filter_key (str): The filter key to use for filtered responses.
Baber's avatar
Baber committed
1798
            indices (dict, optional): Dictionary of sample indices to evaluate.
1799
1800
1801
1802
1803
1804
1805
            rank (int): The process rank.
            limit (int, optional): Limit on number of examples to evaluate.
            world_size (int): Total number of processes.

        Returns:
            list: A list of metrics calculated for each document.
        """
Baber's avatar
Baber committed
1806
1807
        if not requests and not self.instances:
            return None, None
Baber's avatar
rename  
Baber committed
1808
1809
        else:
            requests = requests if requests else self.instances
Baber's avatar
Baber committed
1810

Baber's avatar
Baber committed
1811
1812
1813
        all_metrics = defaultdict(list)
        samples = [] if log_samples else None

Baber's avatar
Baber committed
1814
1815
1816
        ### Collect values of metrics on all datapoints ###
        # Pre-process task.instances to group by doc_id
        instances_by_doc_id = defaultdict(list)
Baber's avatar
rename  
Baber committed
1817
        for instance in requests:
Baber's avatar
Baber committed
1818
1819
1820
1821
            instances_by_doc_id[instance.doc_id].append(instance)
        # Sort instances within each group
        for instances in instances_by_doc_id.values():
            instances.sort(key=lambda x: x.idx)
1822

Baber's avatar
TODO!  
Baber committed
1823
        if filter_keys is None:
1824
1825
1826
1827
1828
            filter_keys = (
                [x.name for x in self._filters]
                if hasattr(self, "_filters")
                else ["none"]
            )
Baber's avatar
TODO!  
Baber committed
1829
1830
1831
1832
1833
1834
1835
        if isinstance(filter_keys, str):
            filter_keys = [filter_keys]
        for filter_key in filter_keys:
            doc_iterator = self.doc_iterator(
                rank=rank,
                limit=limit,
                world_size=world_size,
Baber's avatar
Baber committed
1836
                samples=indices,
Baber's avatar
TODO!  
Baber committed
1837
            )
1838

Baber's avatar
TODO!  
Baber committed
1839
            for doc_id, doc in doc_iterator:
Baber's avatar
Baber committed
1840
                _sample_metric = defaultdict(list)
Baber's avatar
nit  
Baber committed
1841
                _doc_id_true = indices[doc_id] if indices else doc_id
Baber's avatar
Baber committed
1842
                requests = instances_by_doc_id[_doc_id_true]
Baber's avatar
nit  
Baber committed
1843
                if self.OUTPUT_TYPE != "generate_until":
Baber's avatar
Baber committed
1844
                    # if one doc has multiple instances then calculate metric together
Baber's avatar
Baber committed
1845
1846
1847
1848
1849
1850
1851
1852
1853
1854
                    metrics = [
                        self.process_results(
                            doc,
                            list(
                                itertools.chain.from_iterable(
                                    [req.filtered_resps[filter_key] for req in requests]
                                )
                            ),
                        )
                    ]
Baber's avatar
Baber committed
1855
1856
1857
1858
1859
1860
1861
1862
1863
1864
                else:
                    metrics = [
                        self.process_results(doc, response)
                        for req in requests
                        for response in (
                            req.filtered_resps[filter_key]
                            if isinstance(req.filtered_resps[filter_key], list)
                            else [req.filtered_resps[filter_key]]
                        )
                    ]
Baber's avatar
Baber committed
1865
1866
1867
1868
                for metric in metrics:
                    for k, v in metric.items():
                        _sample_metric[k].append(v)
                if log_samples:
Baber's avatar
Baber committed
1869
                    samples.append(
Baber's avatar
Baber committed
1870
1871
1872
1873
                        create_sample_log(
                            doc=doc,
                            doc_id=_doc_id_true,
                            target=self.doc_to_target(doc),
Baber's avatar
Baber committed
1874
1875
                            requests=tuple(requests),
                            metric_names=tuple(str(x) for x in metrics[0]),
Baber's avatar
Baber committed
1876
                            filter_key=filter_key,
Baber's avatar
Baber committed
1877
                            metrics=tuple(metrics),
Baber's avatar
Baber committed
1878
1879
1880
                        )
                    )
                for metric_name, _score in _sample_metric.items():
Baber's avatar
Baber committed
1881
1882
1883
                    all_metrics[(metric_name, filter_key)].append(_score)
        self.metric_results = all_metrics
        return all_metrics, samples
Baber's avatar
test  
Baber committed
1884

Baber's avatar
Baber committed
1885
1886
    def compute_agg_metrics(
        self,
Baber's avatar
Baber committed
1887
        metric_results: dict[tuple[str, str], list[list[float]]] = None,
Baber's avatar
Baber committed
1888
1889
        bootstrap_iters: int = 1000,
    ):
Baber's avatar
Baber committed
1890
        metric_results = metric_results if metric_results else self.metric_results
Baber's avatar
Baber committed
1891
1892
1893
1894
1895
1896
1897
1898
1899
1900
1901
1902
1903
1904
1905
1906
1907
1908
1909
1910
1911
1912
        agg_metrics = defaultdict(list)
        for (metric_name, filter_key), scores in metric_results.items():
            agg_fn = self.aggregation()[metric_name]
            metric_key = f"{metric_name},{filter_key}"
            self.repeat_metric = pass_at_k
            repeats = [
                self.repeat_metric(len(x), x.count(1), k=x.count(1) - 1) for x in scores
            ]
            repeat_agg = np.mean(repeats)
            agg_metrics[metric_key] = [agg_fn(items) for items in zip(*scores)]
            if isinstance(bootstrap_iters, int):
                stderr_fn = stderr_for_metric(
                    metric=agg_fn,
                    bootstrap_iters=min(bootstrap_iters, 100)
                    if metric_name in ["bleu", "chrf", "ter"]
                    else bootstrap_iters,
                )
                agg_metrics[f"{metric_name}_stderr,{filter_key}"] = [
                    (stderr_fn(item) if (stderr_fn and len(item) > 1) else "N/A")
                    for item in zip(*scores)
                ][0]
            agg_metrics[f"{metric_key}_repeat"] = [repeat_agg]
Baber's avatar
test  
Baber committed
1913

Baber's avatar
Baber committed
1914
        return agg_metrics
Baber's avatar
test  
Baber committed
1915

1916
1917

class MultipleChoiceTask(Task):
1918
    OUTPUT_TYPE = "loglikelihood"
1919

baberabb's avatar
baberabb committed
1920
    def doc_to_target(self, doc: dict) -> str:
1921
1922
        return " " + doc["choices"][doc["gold"]]

baberabb's avatar
baberabb committed
1923
    def construct_requests(self, doc: dict, ctx: str, **kwargs) -> List[Instance]:
1924
        # TODO: add mutual info here?
lintangsutawika's avatar
lintangsutawika committed
1925
1926
        return [
            Instance(
haileyschoelkopf's avatar
haileyschoelkopf committed
1927
                request_type="loglikelihood",
lintangsutawika's avatar
lintangsutawika committed
1928
                doc=doc,
1929
                arguments=(ctx, " {}".format(choice)),
1930
                idx=i,
1931
1932
                **kwargs,
            )
lintangsutawika's avatar
lintangsutawika committed
1933
1934
            for i, choice in enumerate(doc["choices"])
        ]
1935

1936
    def process_results(self, doc: dict, results: Iterable[Tuple[float, bool]]) -> dict:
lintangsutawika's avatar
lintangsutawika committed
1937
1938
1939
        results = [
            res[0] for res in results
        ]  # only retain loglikelihoods, discard is_greedy TODO: do we need is_greedy anywhere?
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
        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
1951
    def higher_is_better(self) -> dict:
1952
1953
1954
1955
1956
        return {
            "acc": True,
            "acc_norm": True,
        }

baberabb's avatar
baberabb committed
1957
    def aggregation(self) -> dict:
1958
1959
1960
1961
1962
1963
        return {
            "acc": mean,
            "acc_norm": mean,
        }


lintangsutawika's avatar
lintangsutawika committed
1964
class PerplexityTask(Task):
Baber's avatar
Baber committed
1965
    OUTPUT_TYPE: OutputType = "loglikelihood_rolling"
1966

baberabb's avatar
baberabb committed
1967
    def has_training_docs(self) -> bool:
1968
1969
        return False

baberabb's avatar
baberabb committed
1970
    def fewshot_examples(self, k: int, rnd) -> List:
1971
1972
1973
1974
        if k != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1975
1976
        return []

baberabb's avatar
baberabb committed
1977
    def fewshot_context(self, doc: dict, num_fewshot: int) -> Literal[""]:
1978
1979
1980
1981
        if num_fewshot != 0:
            raise ValueError(
                "The number of fewshot examples must be 0 for perplexity tasks."
            )
1982
1983
1984

        return ""

baberabb's avatar
baberabb committed
1985
    def higher_is_better(self) -> dict:
1986
1987
1988
1989
1990
1991
1992
1993
1994
        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
1995
    def doc_to_text(self, doc) -> str:
1996
1997
1998
1999
2000
        return ""

    def doc_to_target(self, doc):
        return doc

2001
2002
2003
    def construct_requests(self, doc: dict, ctx: Optional[str], **kwargs):
        if bool(ctx):
            raise ValueError
2004

lintangsutawika's avatar
lintangsutawika committed
2005
2006
2007
2008
2009
2010
2011
        return Instance(
            request_type=self.OUTPUT_TYPE,
            doc=doc,
            arguments=(self.doc_to_target(doc),),
            idx=0,
            **kwargs,
        )
2012

2013
    def process_results(self, doc: dict, results: Tuple[float]) -> dict:
2014
        (loglikelihood,) = results
haileyschoelkopf's avatar
haileyschoelkopf committed
2015
2016
        words = self.count_words(self.doc_to_target(doc))
        bytes_ = self.count_bytes(self.doc_to_target(doc))
2017
2018
2019
2020
2021
2022
        return {
            "word_perplexity": (loglikelihood, words),
            "byte_perplexity": (loglikelihood, bytes_),
            "bits_per_byte": (loglikelihood, bytes_),
        }

baberabb's avatar
baberabb committed
2023
    def aggregation(self) -> dict:
2024
2025
2026
2027
2028
2029
2030
        return {
            "word_perplexity": weighted_perplexity,
            "byte_perplexity": weighted_perplexity,
            "bits_per_byte": bits_per_byte,
        }

    @classmethod
baberabb's avatar
baberabb committed
2031
    def count_bytes(cls, doc) -> int:
2032
2033
2034
        return len(doc.encode("utf-8"))

    @classmethod
baberabb's avatar
baberabb committed
2035
    def count_words(cls, doc) -> int:
2036
        """Downstream tasks with custom word boundaries should override this!"""
Lintang Sutawika's avatar
Lintang Sutawika committed
2037
        return len(re.split(r"\s+", doc))