evaluator.py 28.7 KB
Newer Older
Baber Abbasi's avatar
Baber Abbasi committed
1
import itertools
2
import json
3
import logging
Baber Abbasi's avatar
Baber Abbasi committed
4
import random
5
import time
6
7
from collections import defaultdict
from typing import TYPE_CHECKING, List, Optional, Union
Baber Abbasi's avatar
Baber Abbasi committed
8

9
import numpy as np
Baber Abbasi's avatar
Baber Abbasi committed
10
import torch
lintangsutawika's avatar
lintangsutawika committed
11

lintangsutawika's avatar
lintangsutawika committed
12
import lm_eval.api.metrics
lintangsutawika's avatar
lintangsutawika committed
13
import lm_eval.api.registry
Lintang Sutawika's avatar
Lintang Sutawika committed
14
import lm_eval.api.task
Baber Abbasi's avatar
Baber Abbasi committed
15
import lm_eval.models
16
from lm_eval.caching.cache import delete_cache
17
from lm_eval.evaluator_utils import (
Lintang Sutawika's avatar
Lintang Sutawika committed
18
    consolidate_group_results,
19
20
    consolidate_results,
    get_sample_size,
Lintang Sutawika's avatar
Lintang Sutawika committed
21
    get_subtask_list,
22
23
24
25
26
    get_task_list,
    prepare_print_tasks,
    print_writeout,
    run_task_tests,
)
KonradSzafer's avatar
KonradSzafer committed
27
from lm_eval.loggers import EvaluationTracker
28
from lm_eval.loggers.utils import add_env_info, add_tokenizer_info, get_git_commit_hash
Lintang Sutawika's avatar
Lintang Sutawika committed
29
30
31
32
from lm_eval.tasks import (
    TaskManager,
    get_task_dict,
)
33
34
35
36
37
38
39
from lm_eval.utils import (
    eval_logger,
    handle_non_serializable,
    hash_string,
    positional_deprecated,
    simple_parse_args_string,
)
40

Fabrizio Milo's avatar
Fabrizio Milo committed
41

42
43
if TYPE_CHECKING:
    from lm_eval.api.model import LM
Lintang Sutawika's avatar
Lintang Sutawika committed
44
    from lm_eval.api.task import Task
45
46


47
@positional_deprecated
Fabrizio Milo's avatar
Fabrizio Milo committed
48
49
def simple_evaluate(
    model,
50
51
    model_args: Optional[Union[str, dict]] = None,
    tasks: Optional[List[Union[str, dict, object]]] = None,
Baber Abbasi's avatar
Baber Abbasi committed
52
    num_fewshot: Optional[int] = None,
53
    batch_size: Optional[Union[int, str]] = None,
Baber Abbasi's avatar
Baber Abbasi committed
54
55
56
    max_batch_size: Optional[int] = None,
    device: Optional[str] = None,
    use_cache: Optional[str] = None,
57
58
59
    cache_requests: bool = False,
    rewrite_requests_cache: bool = False,
    delete_requests_cache: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
60
    limit: Optional[Union[int, float]] = None,
Ethan Smith's avatar
Ethan Smith committed
61
62
63
64
    bootstrap_iters: int = 100000,
    check_integrity: bool = False,
    write_out: bool = False,
    log_samples: bool = True,
KonradSzafer's avatar
KonradSzafer committed
65
66
    evaluation_tracker: Optional[EvaluationTracker] = None,
    system_instruction: Optional[str] = None,
67
    apply_chat_template: Union[bool, str] = False,
KonradSzafer's avatar
KonradSzafer committed
68
    fewshot_as_multiturn: bool = False,
69
70
    gen_kwargs: Optional[str] = None,
    task_manager: Optional[TaskManager] = None,
71
    verbosity: str = "INFO",
Baber Abbasi's avatar
Baber Abbasi committed
72
    predict_only: bool = False,
73
74
75
    random_seed: int = 0,
    numpy_random_seed: int = 1234,
    torch_random_seed: int = 1234,
76
    fewshot_random_seed: int = 1234,
Hojin Lee's avatar
Hojin Lee committed
77
    confirm_run_unsafe_code: bool = False,
Baber's avatar
Baber committed
78
    mcq_to_generative: bool = False,
Fabrizio Milo's avatar
Fabrizio Milo committed
79
):
80
    """Instantiate and evaluate a model on a list of tasks.
81

82
83
    :param model: Union[str, LM]
        Name of model or LM object, see lm_eval.models.get_model
84
85
    :param model_args: Optional[str, dict]
        String or dict arguments for each model class, see LM.create_from_arg_string and LM.create_from_arg_object.
86
        Ignored if `model` argument is a LM object.
87
    :param tasks: list[Union[str, dict, Task]]
Leo Gao's avatar
Leo Gao committed
88
        List of task names or Task objects. Task objects will be taken to have name task.EVAL_HARNESS_NAME if defined and type(task).__name__ otherwise.
89
90
    :param num_fewshot: int
        Number of examples in few-shot context
91
    :param batch_size: int or str, optional
92
        Batch size for model
93
94
    :param max_batch_size: int, optional
        Maximal batch size to try with automatic batch size detection
95
    :param device: str, optional
96
        PyTorch device (e.g. "cpu" or "cuda:0") for running models
haileyschoelkopf's avatar
haileyschoelkopf committed
97
98
    :param use_cache: str, optional
        A path to a sqlite db file for caching model responses. `None` if not caching.
99
100
101
102
103
104
    :param cache_requests: bool, optional
        Speed up evaluation by caching the building of dataset requests. `None` if not caching.
    :param rewrite_requests_cache: bool, optional
        Rewrites all of the request cache if set to `True`. `None` if not desired.
    :param delete_requests_cache: bool, optional
        Deletes all of the request cache if set to `True`. `None` if not desired.
105
106
    :param limit: int or float, optional
        Limit the number of examples per task (only use this for testing), If <1, limit is a percentage of the total number of examples.
107
    :param bootstrap_iters:
108
        Number of iterations for bootstrap statistics, used when calculating stderrs. set to 0 for no stderr calculations to be performed.
Stephen Hogg's avatar
Stephen Hogg committed
109
110
    :param check_integrity: bool
        Whether to run the relevant part of the test suite for the tasks
111
    :param write_out: bool
112
113
114
        If True, write out an example document and model input for checking task integrity
    :param log_samples: bool
        If True, write out all model outputs and documents for per-sample measurement and post-hoc analysis
KonradSzafer's avatar
KonradSzafer committed
115
116
    :param system_instruction: str
        System instruction to be applied to the prompt
117
118
119
120
121
    :param apply_chat_template: Union[bool, str]
        Specifies whether to apply a chat template to the prompt.
        - If set to True, the default chat template is applied.
        - If set to a string, applies the specified chat template by name.
        Defaults to False (no chat template applied).
KonradSzafer's avatar
KonradSzafer committed
122
123
    :param fewshot_as_multiturn: bool
        Whether to provide the fewshot examples as a multiturn conversation or a single user turn.
124
125
126
    :param gen_kwargs: str
        String arguments for model generation
        Ignored for all tasks with loglikelihood output_type
Baber Abbasi's avatar
Baber Abbasi committed
127
128
    :param predict_only: bool
        If true only model outputs will be generated and returned. Metrics will not be evaluated
129
130
131
132
133
134
    :param random_seed: int
        Random seed for python's random module. If set to None, the seed will not be set.
    :param numpy_random_seed: int
        Random seed for numpy. If set to None, the seed will not be set.
    :param torch_random_seed: int
        Random seed for torch. If set to None, the seed will not be set.
135
136
    :param fewshot_random_seed: int
        Random seed for fewshot sampler random generator. If set to None, the seed of generator will be set to None.
Baber Abbasi's avatar
Baber Abbasi committed
137

138
    :return
139
        Dictionary of results
140
    """
141
    eval_logger.setLevel(getattr(logging, f"{verbosity}"))
142
    start_date = time.time()
143

144
145
146
147
    if delete_requests_cache:
        eval_logger.info("Deleting requests cache...")
        delete_cache()

148
    seed_message = []
149
150
    if random_seed is not None:
        # See https://github.com/EleutherAI/lm-evaluation-harness/pull/1412
151
        seed_message.append(f"Setting random seed to {random_seed}")
152
153
154
        random.seed(random_seed)

    if numpy_random_seed is not None:
155
        seed_message.append(f"Setting numpy seed to {numpy_random_seed}")
156
157
158
        np.random.seed(numpy_random_seed)

    if torch_random_seed is not None:
159
        seed_message.append(f"Setting torch manual seed to {torch_random_seed}")
160
161
        torch.manual_seed(torch_random_seed)

162
163
164
    if fewshot_random_seed is not None:
        seed_message.append(f"Setting fewshot manual seed to {fewshot_random_seed}")

165
166
167
    if seed_message:
        eval_logger.info(" | ".join(seed_message))

168
169
    if tasks is None:
        tasks = []
170
171
172
173
    if len(tasks) == 0:
        raise ValueError(
            "No tasks specified, or no tasks found. Please verify the task names."
        )
174

lintangsutawika's avatar
lintangsutawika committed
175
176
    if gen_kwargs is not None:
        gen_kwargs = simple_parse_args_string(gen_kwargs)
lintangsutawika's avatar
udate  
lintangsutawika committed
177
        eval_logger.warning(
178
179
            "generation_kwargs specified through cli, these settings will update set parameters in yaml tasks. "
            "Ensure 'do_sample=True' for non-greedy decoding!"
lintangsutawika's avatar
udate  
lintangsutawika committed
180
        )
lintangsutawika's avatar
lintangsutawika committed
181
182
183
        if gen_kwargs == "":
            gen_kwargs = None

184
    if isinstance(model, str):
Fabrizio Milo's avatar
Fabrizio Milo committed
185
        if model_args is None:
186
            eval_logger.warning("model_args not specified. Using defaults.")
Fabrizio Milo's avatar
Fabrizio Milo committed
187
            model_args = ""
188

189
        if isinstance(model_args, dict):
190
191
192
            eval_logger.info(
                f"Initializing {model} model, with arguments: {model_args}"
            )
193
194
195
196
197
198
199
200
201
202
            lm = lm_eval.api.registry.get_model(model).create_from_arg_obj(
                model_args,
                {
                    "batch_size": batch_size,
                    "max_batch_size": max_batch_size,
                    "device": device,
                },
            )

        else:
203
204
205
            eval_logger.info(
                f"Initializing {model} model, with arguments: {simple_parse_args_string(model_args)}"
            )
206
207
208
209
210
211
212
213
            lm = lm_eval.api.registry.get_model(model).create_from_arg_string(
                model_args,
                {
                    "batch_size": batch_size,
                    "max_batch_size": max_batch_size,
                    "device": device,
                },
            )
214
    else:
215
        if not isinstance(model, lm_eval.api.model.LM):
216
217
218
            raise TypeError(
                f"The value of `model` passed to simple_evaluate() was of type {type(model)}, but is required to be a subclass of lm_eval.api.model.LM . This may be because you are passing an initialized Hugging Face PreTrainedModel without having wrapped it in `lm_eval.models.huggingface.HFLM(pretrained=my_model)` first."
            )
219
        eval_logger.info("Using pre-initialized model")
220
        lm = model
221

haileyschoelkopf's avatar
haileyschoelkopf committed
222
    if use_cache is not None:
223
        eval_logger.info(f"Using cache at {use_cache + '_rank' + str(lm.rank) + '.db'}")
haileyschoelkopf's avatar
haileyschoelkopf committed
224
225
226
227
228
        lm = lm_eval.api.model.CachingLM(
            lm,
            use_cache
            # each rank receives a different cache db.
            # necessary to avoid multiple writes to cache at once
229
230
231
            + "_rank"
            + str(lm.rank)
            + ".db",
haileyschoelkopf's avatar
haileyschoelkopf committed
232
233
        )

234
    if task_manager is None:
Baber's avatar
Baber committed
235
        task_manager = TaskManager(verbosity, mcq_to_generative=mcq_to_generative)
236
237

    task_dict = get_task_dict(tasks, task_manager)
Baber Abbasi's avatar
Baber Abbasi committed
238

Lintang Sutawika's avatar
Lintang Sutawika committed
239
240
241
242
243
244
245
246
247
248
    # helper function to recursively apply config overrides to leaf subtasks, skipping their constituent groups.
    # (setting of num_fewshot ; bypassing metric calculation ; setting fewshot seed)
    def _adjust_config(task_dict):
        adjusted_task_dict = {}
        for task_name, task_obj in task_dict.items():
            if isinstance(task_obj, dict):
                adjusted_task_dict = {
                    **adjusted_task_dict,
                    **{task_name: _adjust_config(task_obj)},
                }
249

250
            else:
Lintang Sutawika's avatar
Lintang Sutawika committed
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
                if task_obj.get_config("output_type") == "generate_until":
                    if gen_kwargs is not None:
                        task_obj.set_config(
                            key="generation_kwargs", value=gen_kwargs, update=True
                        )

                if predict_only:
                    eval_logger.info(
                        f"Processing {task_name} in output-only mode. Metrics will not be calculated!"
                    )
                    # we have to change the class properties post-hoc. This is pretty hacky.
                    task_obj.override_metric(metric_name="bypass")

                # override tasks' fewshot values to the provided num_fewshot arg value
                # except if tasks have it set to 0 manually in their configs--then we should never overwrite that
                if num_fewshot is not None:
                    if (default_num_fewshot := task_obj.get_config("num_fewshot")) == 0:
                        eval_logger.info(
                            f"num_fewshot has been set to 0 for {task_name} in its config. Manual configuration will be ignored."
                        )
                    else:
                        eval_logger.warning(
                            f"Overwriting default num_fewshot of {task_name} from {default_num_fewshot} to {num_fewshot}"
                        )
                        task_obj.set_config(key="num_fewshot", value=num_fewshot)
                else:
                    # if num_fewshot not provided, and the task does not define a default one, default to 0
                    if (
                        default_num_fewshot := task_obj.get_config("num_fewshot")
                    ) is None:
                        task_obj.set_config(key="num_fewshot", value=0)
                # fewshot_random_seed set for tasks, even with a default num_fewshot (e.g. in the YAML file)
                task_obj.set_fewshot_seed(seed=fewshot_random_seed)

                adjusted_task_dict[task_name] = task_obj

        return adjusted_task_dict

    task_dict = _adjust_config(task_dict)
Jonathan Tow's avatar
Merge  
Jonathan Tow committed
290

Stephen Hogg's avatar
Stephen Hogg committed
291
    if check_integrity:
292
        run_task_tests(task_list=tasks)
Stephen Hogg's avatar
Stephen Hogg committed
293

KonradSzafer's avatar
KonradSzafer committed
294
295
296
297
298
    if evaluation_tracker is not None:
        evaluation_tracker.general_config_tracker.log_experiment_args(
            model_source=model,
            model_args=model_args,
            system_instruction=system_instruction,
Baber Abbasi's avatar
Baber Abbasi committed
299
300
301
            chat_template=lm.chat_template(apply_chat_template)
            if apply_chat_template
            else None,
302
            fewshot_as_multiturn=fewshot_as_multiturn,
KonradSzafer's avatar
KonradSzafer committed
303
304
        )

305
306
307
308
    results = evaluate(
        lm=lm,
        task_dict=task_dict,
        limit=limit,
309
310
        cache_requests=cache_requests,
        rewrite_requests_cache=rewrite_requests_cache,
Niklas Muennighoff's avatar
Niklas Muennighoff committed
311
        bootstrap_iters=bootstrap_iters,
312
        write_out=write_out,
Lintang Sutawika's avatar
Lintang Sutawika committed
313
        log_samples=True if predict_only else log_samples,
KonradSzafer's avatar
KonradSzafer committed
314
315
316
        system_instruction=system_instruction,
        apply_chat_template=apply_chat_template,
        fewshot_as_multiturn=fewshot_as_multiturn,
317
        verbosity=verbosity,
Hojin Lee's avatar
Hojin Lee committed
318
        confirm_run_unsafe_code=confirm_run_unsafe_code,
319
    )
320

321
    if lm.rank == 0:
322
323
324
325
326
327
328
        if isinstance(model, str):
            model_name = model
        elif hasattr(model, "config") and hasattr(model.config, "_name_or_path"):
            model_name = model.config._name_or_path
        else:
            model_name = type(model).__name__

329
330
        # add info about the model and few shot config
        results["config"] = {
331
            "model": model_name,
332
333
            "model_args": model_args,
        }
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
        # add more detailed model info if available
        if isinstance(lm, lm_eval.models.huggingface.HFLM):
            results["config"].update(lm.get_model_info())
        # add info about execution
        results["config"].update(
            {
                "batch_size": batch_size,
                "batch_sizes": (
                    list(lm.batch_sizes.values()) if hasattr(lm, "batch_sizes") else []
                ),
                "device": device,
                "use_cache": use_cache,
                "limit": limit,
                "bootstrap_iters": bootstrap_iters,
                "gen_kwargs": gen_kwargs,
349
350
351
352
                "random_seed": random_seed,
                "numpy_seed": numpy_random_seed,
                "torch_seed": torch_random_seed,
                "fewshot_seed": fewshot_random_seed,
353
354
            }
        )
355
        results["git_hash"] = get_git_commit_hash()
356
        results["date"] = start_date
357
        add_env_info(results)  # additional environment info to results
achervyakov's avatar
achervyakov committed
358
        add_tokenizer_info(results, lm)  # additional info about tokenizer
359
360
361
        return results
    else:
        return None
362

Leo Gao's avatar
Leo Gao committed
363

364
@positional_deprecated
Fabrizio Milo's avatar
Fabrizio Milo committed
365
def evaluate(
366
    lm: "LM",
Fabrizio Milo's avatar
Fabrizio Milo committed
367
    task_dict,
Baber Abbasi's avatar
Baber Abbasi committed
368
    limit: Optional[int] = None,
369
370
    cache_requests: bool = False,
    rewrite_requests_cache: bool = False,
Baber Abbasi's avatar
Baber Abbasi committed
371
    bootstrap_iters: Optional[int] = 100000,
Ethan Smith's avatar
Ethan Smith committed
372
373
    write_out: bool = False,
    log_samples: bool = True,
KonradSzafer's avatar
KonradSzafer committed
374
    system_instruction: Optional[str] = None,
375
    apply_chat_template: Union[bool, str] = False,
KonradSzafer's avatar
KonradSzafer committed
376
    fewshot_as_multiturn: bool = False,
377
    verbosity: str = "INFO",
Hojin Lee's avatar
Hojin Lee committed
378
    confirm_run_unsafe_code: bool = False,
Fabrizio Milo's avatar
Fabrizio Milo committed
379
):
380
381
382
383
384
    """Instantiate and evaluate a model on a list of tasks.

    :param lm: obj
        Language Model
    :param task_dict: dict[str, Task]
haileyschoelkopf's avatar
haileyschoelkopf committed
385
        Dictionary of tasks. Tasks will be taken to have name type(task).config.task .
386
387
    :param limit: int, optional
        Limit the number of examples per task (only use this for testing)
Hojin Lee's avatar
Hojin Lee committed
388
389
390
391
    :param cache_requests: bool, optional
        Speed up evaluation by caching the building of dataset requests.
    :param rewrite_requests_cache: bool, optional
        Rewrites all the request cache if set to `True`.
392
    :param bootstrap_iters:
393
        Number of iterations for bootstrap statistics, used when calculating stderr. Set to 0 for skipping all stderr calculations.
394
    :param write_out: bool
395
396
397
        If True, write out an example document and model input for checking task integrity
    :param log_samples: bool
        If True, write out all model outputs and documents for per-sample measurement and post-hoc analysis
KonradSzafer's avatar
KonradSzafer committed
398
399
    :param system_instruction: str
        System instruction to be applied to the prompt
400
401
402
403
404
    :param apply_chat_template: Union[bool, str]
        Specifies whether to apply a chat template to the prompt.
        - If set to True, the default chat template is applied.
        - If set to a string, applies the specified chat template by name.
        Defaults to False (no chat template applied).
KonradSzafer's avatar
KonradSzafer committed
405
406
    :param fewshot_as_multiturn: bool
        Whether to provide the fewshot examples as a multiturn conversation or a single user turn.
Hojin Lee's avatar
Hojin Lee committed
407
408
409
410
    :param verbosity: str
        Verbosity level for logging
    :param confirm_run_unsafe_code: bool
        Whether to confirm running tasks marked as unsafe.
411
412
413
    :return
        Dictionary of results
    """
414

415
    eval_logger.setLevel(getattr(logging, f"{verbosity}"))
416

417
418
419
420
421
    if apply_chat_template:
        eval_logger.warning(
            "Chat template formatting change affects loglikelihood and multiple-choice tasks. See docs/chat-template-readme.md for details."
        )

422
    # tracks all Instances/requests a model must generate output on.
423
    requests = defaultdict(list)
424
425
    # stores the amount to pad out reqs per req. type so that
    # number of fwd passes per distributed rank is equal
426
    padding_requests = defaultdict(int)
427

428
    # get lists of group hierarchy and each type of request
Lintang Sutawika's avatar
Lintang Sutawika committed
429
    eval_tasks = get_task_list(task_dict)
430
    if not log_samples:
431
        if not all(
432
433
            "bypass" not in getattr(task_output.task, "_metric_fn_list", {}).keys()
            for task_output in eval_tasks
434
435
        ):
            raise ValueError("log_samples must be True for 'bypass' metric-only tasks")
436

Hojin Lee's avatar
Hojin Lee committed
437
438
439
    # validation checks:
    # 1.are we running multimodal task <-> non-multimodal model class, or vice-versa.
    # 2.are we running code that is marked as unsafe.
440
    incompatible_tasks = []
441
442
    for task_output in eval_tasks:
        task: Task = task_output.task
443
444
445

        if getattr(lm, "MULTIMODAL", False) != getattr(task, "MULTIMODAL", False):
            incompatible_tasks.append(task_output.task_name)
Hojin Lee's avatar
Hojin Lee committed
446
447
448
449
        elif getattr(task, "UNSAFE_CODE", False) and not confirm_run_unsafe_code:
            raise ValueError(
                f"Attempted to run task: {task_output.task_name} which is marked as unsafe. Set confirm_run_unsafe_code=True to run this task."
            )
450
451
452
453
454
455
456
457
458
    if len(incompatible_tasks) > 0:
        if not getattr(lm, "MULTIMODAL", False):
            raise ValueError(
                f"Attempted to run tasks: {incompatible_tasks} which require multimodal input, but the selected model type does not currently implement this. Multimodal support is currently restricted to the ['hf-multimodal', 'vllm-vlm'] model type."
            )
        else:
            raise ValueError(
                f"Attempted to run tasks: {incompatible_tasks} which are text-only, but used a model type which only currently supports multimodal tasks."
            )
Hojin Lee's avatar
Hojin Lee committed
459
    # end validation check
460

Chenjie Luo's avatar
Chenjie Luo committed
461
462
463
    # Cache the limit arg.
    limit_arg = limit
    limits = []
464
465
466
    for task_output in eval_tasks:
        task: Task = task_output.task

Chenjie Luo's avatar
Chenjie Luo committed
467
468
        limit = get_sample_size(task, limit_arg)
        limits.append(limit)
469
470
471
472
473
474
        task.build_all_requests(
            limit=limit,
            rank=lm.rank,
            world_size=lm.world_size,
            cache_requests=cache_requests,
            rewrite_requests_cache=rewrite_requests_cache,
KonradSzafer's avatar
KonradSzafer committed
475
            system_instruction=system_instruction,
476
            apply_chat_template=bool(apply_chat_template),
KonradSzafer's avatar
KonradSzafer committed
477
            fewshot_as_multiturn=fewshot_as_multiturn,
478
479
480
481
482
483
            chat_template=getattr(lm, "apply_chat_template")
            if apply_chat_template
            else None,
            tokenizer_name=getattr(lm, "tokenizer_name", "")
            if apply_chat_template
            else "",
484
        )
485
        eval_logger.debug(
486
            f"Task: {task_output.task_name}; number of requests on this rank: {len(task.instances)}"
haileyschoelkopf's avatar
haileyschoelkopf committed
487
488
        )
        if write_out:
489
            print_writeout(task)
490
        # aggregate Instances by LM method requested to get output.
lintangsutawika's avatar
lintangsutawika committed
491
492
493
        for instance in task.instances:
            reqtype = instance.request_type
            requests[reqtype].append(instance)
494
495

        if lm.world_size > 1:
496
497
498
499
            instances_rnk = torch.tensor(len(task._instances), device=lm.device)
            gathered_item = (
                lm.accelerator.gather(instances_rnk).cpu().detach().numpy().tolist()
            )
500
501
502
503
504
505
            # "multiple_choice" task types dispatch (several) "loglikelihood" request types
            reqtype = (
                "loglikelihood"
                if task.OUTPUT_TYPE == "multiple_choice"
                else task.OUTPUT_TYPE
            )
506
            # compute number of pseudo-batches to pad with (FSDP/DDP require even batches among ranks)
507
            numpad = max(gathered_item) - gathered_item[lm.rank]
508
509
            # todo: may not account for padding in cases like SquadV2 which has multiple req types
            padding_requests[reqtype] += numpad
510

511
    ### Run LM on inputs, get all outputs ###
Leo Gao's avatar
Leo Gao committed
512
513
    # execute each type of request
    for reqtype, reqs in requests.items():
514
        eval_logger.info(f"Running {reqtype} requests")
515
516
517
518
        # create `K` copies of each request `req` based off `K = req.repeats`
        cloned_reqs = []
        for req in reqs:
            cloned_reqs.extend([req] * req.repeats)
lintangsutawika's avatar
lintangsutawika committed
519

520
521
        if (lm.world_size > 1) and (padding_requests[reqtype] > 0):
            for _ in range(padding_requests[reqtype]):
522
523
                cloned_reqs.extend([req] * req.repeats)

524
525
526
527
528
529
530
        # run requests through model
        resps = getattr(lm, reqtype)(cloned_reqs)

        # put responses from model into a list of length K for each request.
        for x, req in zip(resps, cloned_reqs):
            req.resps.append(x)

531
532
        if lm.world_size > 1:
            lm.accelerator.wait_for_everyone()
533

534
535
    RANK = lm.rank
    WORLD_SIZE = lm.world_size
536
537
    ### Postprocess outputs ###
    # TODO: del model here, maybe (idea: allow user to specify device of e.g. reward model separately)
Chenjie Luo's avatar
Chenjie Luo committed
538
    for task_output, limit in zip(eval_tasks, limits):
539
        task = task_output.task
540
541
        task.apply_filters()

542
543
        ### Collect values of metrics on all datapoints ###
        # # unpack results and sort back in order and return control to Task
haileyschoelkopf's avatar
haileyschoelkopf committed
544
        # TODO: make it possible to use a different metric per filter
545
        # Pre-process task.instances to group by doc_id
546
        instances_by_doc_id = defaultdict(list)
547
548
549
550
551
        for instance in task.instances:
            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)
haileyschoelkopf's avatar
haileyschoelkopf committed
552
        # iterate over different filters used
553
554
555
        for filter_key in task.instances[0].filtered_resps.keys():
            doc_iterator = task.doc_iterator(
                rank=RANK, limit=limit, world_size=WORLD_SIZE
556
            )
557
            for doc_id, doc in doc_iterator:
558
                requests = instances_by_doc_id[doc_id]
lintangsutawika's avatar
lintangsutawika committed
559
                metrics = task.process_results(
560
                    doc, [req.filtered_resps[filter_key] for req in requests]
lintangsutawika's avatar
lintangsutawika committed
561
                )
562
563
564
565
566
567
568
569
                if log_samples:
                    target = task.doc_to_target(doc)
                    example = {
                        "doc_id": doc_id,
                        "doc": doc,
                        "target": target,
                        "arguments": [req.args for req in requests],
                        "resps": [req.resps for req in requests],
570
571
572
                        "filtered_resps": [
                            req.filtered_resps[filter_key] for req in requests
                        ],
573
574
                        "filter": filter_key,
                        "metrics": list(metrics.keys()),
575
576
577
578
579
580
581
582
583
584
                        "doc_hash": hash_string(
                            json.dumps(
                                requests[0].doc,
                                indent=2,
                                default=handle_non_serializable,
                                ensure_ascii=False,
                            )
                        ),
                        "prompt_hash": hash_string(requests[0].arguments[0]),
                        "target_hash": hash_string(str(target)),
585
586
                    }
                    example.update(metrics)
587
                    task_output.logged_samples.append(example)
588
                for metric, value in metrics.items():
589
                    task_output.sample_metrics[(metric, filter_key)].append(value)
590

591
592
    if WORLD_SIZE > 1:
        # if multigpu, then gather data across all ranks to rank 0
593
        # first gather logged samples across all ranks
594
595
596
597
598
599
600
601
        for task_output in eval_tasks:
            if log_samples:
                # for task_name, task_samples in list(samples.items()):
                full_samples = [None] * WORLD_SIZE if RANK == 0 else None
                torch.distributed.gather_object(
                    obj=task_output.logged_samples,
                    object_gather_list=full_samples,
                    dst=0,
602
                )
603

604
605
606
607
                if RANK == 0:
                    task_output.logged_samples = list(
                        itertools.chain.from_iterable(full_samples)
                    )
608

609
610
611
612
613
614
615
            # then collect metrics across all ranks
            for metrics in task_output.sample_metrics:
                metric_list = [None] * WORLD_SIZE if RANK == 0 else None
                torch.distributed.gather_object(
                    obj=task_output.sample_metrics[metrics],
                    object_gather_list=metric_list,
                    dst=0,
616
                )
617
618
619
620
                if RANK == 0:
                    task_output.sample_metrics[metrics] = list(
                        itertools.chain.from_iterable(metric_list)
                    )
621

622
    if RANK == 0:
623
624
        ### Aggregate results over all datapoints ###
        # aggregate results ; run bootstrap CIs
625
626
        for task_output in eval_tasks:
            task_output.calculate_aggregate_metric(bootstrap_iters=bootstrap_iters)
627
628
629
630
631
632
633
634
        (
            results,
            samples,
            configs,
            versions,
            num_fewshot,
            higher_is_better,
        ) = consolidate_results(eval_tasks)
Fabrizio Milo's avatar
Fabrizio Milo committed
635

636
        ### Calculate group metrics ###
lintangsutawika's avatar
lintangsutawika committed
637
        if bool(results):
Lintang Sutawika's avatar
Lintang Sutawika committed
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
            results, versions, show_group_table, *_ = consolidate_group_results(
                results, versions, task_dict
            )

        results_agg, group_agg = prepare_print_tasks(task_dict, results)
        subtask_list = get_subtask_list(task_dict)

        # collect all higher_is_better values for metrics
        # in the group's subtasks.
        # TODO: clean this up ; unify with the below metric_list loop?
        _higher_is_better = {}
        for group, task_list in subtask_list.items():
            if (
                len(task_list) != 0
            ):  # subtask list will list "task_name": [] for solo tasks
653
654
655
656
                for task in task_list:
                    for m, h in higher_is_better[task].items():
                        if m not in _higher_is_better.keys():
                            _higher_is_better[m] = h
lintangsutawika's avatar
lintangsutawika committed
657

Lintang Sutawika's avatar
Lintang Sutawika committed
658
659
660
661
662
663
664
665
666
667
                        if (
                            m in _higher_is_better
                            and _higher_is_better[m] is not None
                            and _higher_is_better[m] != h
                        ):
                            eval_logger.warning(
                                f"Higher_is_better values for metric {m} in group {group} are not consistent. Defaulting to None."
                            )
                            _higher_is_better[m] = None
                higher_is_better[group] = _higher_is_better
668

669
        results_dict = {
670
            "results": dict(results_agg.items()),
Lintang Sutawika's avatar
Lintang Sutawika committed
671
672
673
674
675
676
            **(
                {"groups": dict(group_agg.items())}
                if (bool(group_agg) & show_group_table)
                else {}
            ),
            "group_subtasks": dict(reversed(subtask_list.items())),
677
678
            "configs": dict(sorted(configs.items())),
            "versions": dict(sorted(versions.items())),
679
            "n-shot": dict(sorted(num_fewshot.items())),
680
            "higher_is_better": dict(sorted(higher_is_better.items())),
681
682
683
            "n-samples": {
                task_output.task_name: {
                    "original": len(task_output.task.eval_docs),
KonradSzafer's avatar
KonradSzafer committed
684
685
686
687
                    "effective": min(
                        limit if limit else len(task_output.task.eval_docs),
                        len(task_output.task.eval_docs),
                    ),
688
                }
Chenjie Luo's avatar
Chenjie Luo committed
689
                for task_output, limit in zip(eval_tasks, limits)
690
            },
691
        }
692
693
694
695
        if log_samples:
            results_dict["samples"] = dict(samples)

        return results_dict
Fabrizio Milo's avatar
Fabrizio Milo committed
696

697
698
    else:
        return None
699
700
701
702


def request_caching_arg_to_dict(cache_requests: str) -> dict:
    request_caching_args = {
703
704
705
        "cache_requests": cache_requests in {"true", "refresh"},
        "rewrite_requests_cache": cache_requests == "refresh",
        "delete_requests_cache": cache_requests == "delete",
706
707
708
    }

    return request_caching_args