merged with conflict resolved

.gitignore

@@ -3,3 +3,5 @@ env
 data/
 lm_cache
 .idea
+
+*.egg-info/

.pre-commit-config.yaml

@@ -12,6 +12,7 @@ repos:
       - id: check-merge-conflict
       - id: check-symlinks
       - id: check-yaml
+        args: ['--unsafe']
       - id: destroyed-symlinks
       - id: detect-private-key
       - id: end-of-file-fixer

README.md

@@ -45,7 +45,7 @@ python main.py \
     --device cuda:0
 ```
 
-Additional arguments can be provided to the model constructor using the `--model_args` flag. Most notably, this supports the common practice of using the `revisions` feature on the Hub to store partialy trained checkpoints:
+Additional arguments can be provided to the model constructor using the `--model_args` flag. Most notably, this supports the common practice of using the `revisions` feature on the Hub to store partially trained checkpoints:
 
 ```bash
 python main.py \
@@ -64,8 +64,8 @@ To use with [PEFT](https://github.com/huggingface/peft), take the call you would
 python main.py \
     --model hf-causal \
     --model_args pretrained=EleutherAI/gpt-j-6b,peft=nomic-ai/gpt4all-j-lora \
-    --tasks openbookqa,arc_easy,winogrande,hellaswag,arc_challenge,piqa,boolq \ 
-    --device cuda:0 
+    --tasks openbookqa,arc_easy,winogrande,hellaswag,arc_challenge,piqa,boolq \
+    --device cuda:0
 ```
 
 Our library also supports the OpenAI API:
@@ -78,7 +78,7 @@ python main.py \
     --tasks lambada_openai,hellaswag
 ```
 
-While this functionality is only officially mantained for the official OpenAI API, it tends to also work for other hosting services that use the same API such as [goose.ai](goose.ai) with minor modification. We also have an implementation for the [TextSynth](https://textsynth.com/index.html) API, using `--model textsynth`.
+While this functionality is only officially maintained for the official OpenAI API, it tends to also work for other hosting services that use the same API such as [goose.ai](goose.ai) with minor modification. We also have an implementation for the [TextSynth](https://textsynth.com/index.html) API, using `--model textsynth`.
 
 To verify the data integrity of the tasks you're performing in addition to running the tasks themselves, you can use the `--check_integrity` flag:
 
@@ -116,7 +116,7 @@ When reporting eval harness results, please also report the version of each task
 
 ## Test Set Decontamination
 
-To address concerns about train / test contamination, we provide utilities for comparing results on a benchmark using only the data points nto found in the model trainign set. Unfortunately, outside of models trained on the Pile ans C4, its very rare that people who train models disclose the contents of the training data. However this utility can be useful to evaluate models you have trained on private data, provided you are willing to pre-compute the necessary indices. We provide computed indices for 13-gram exact match deduplication against the Pile, and plan to add additional precomputed dataset indices in the future (including C4 and min-hash LSH deduplication).
+To address concerns about train / test contamination, we provide utilities for comparing results on a benchmark using only the data points nto found in the model trainign set. Unfortunately, outside of models trained on the Pile and C4, its very rare that people who train models disclose the contents of the training data. However this utility can be useful to evaluate models you have trained on private data, provided you are willing to pre-compute the necessary indices. We provide computed indices for 13-gram exact match deduplication against the Pile, and plan to add additional precomputed dataset indices in the future (including C4 and min-hash LSH deduplication).
 
 For details on text decontamination, see the [decontamination guide](./docs/decontamination.md).
 

lm_eval/api/__init__.py

 from . import metrics
-
-METRIC_REGISTRY = {
-    "matthews_corrcoef": metrics.matthews_corrcoef,
-    "f1_score": metrics.f1_score,
-    "perplexity": metrics.perplexity,
-    "bleu": metrics.bleu,
-    "chrf": metrics.chrf,
-    "ter": metrics.ter,
-}
-
-AGGREGATION_REGISTRY = {
-    "mean": metrics.mean,
-    "median": metrics.median,
-    "perplexity": metrics.perplexity,
-}
-
-HIGHER_IS_BETTER_REGISTRY = {
-    "matthews_corrcoef": True,
-    "f1_score": True,
-    "perplexity": False,
-    "bleu": True,
-    "chrf": True,
-    "ter": False,
-
-    "acc": True,
-    "acc_norm": True,
-    "word_perplexity": False,
-    "byte_perplexity": False,
-    "bits_per_byte": False,
-}
\ No newline at end of file

lm_eval/api/filter.py

@@ -3,9 +3,10 @@ from typing import List
 
 from lm_eval.api.instance import Instance
 
+
 class Filter:
     """
-    Filter classes operate on a per-task level. 
+    Filter classes operate on a per-task level.
     They take all model outputs (`instance.resps` for all `task.instances`)
     across all instances of a task, and perform operations.
     In a single run, one can configure any number of separate filters or lists of filters.
@@ -25,30 +26,33 @@ class Filter:
         [<filtered resps for instance 0>, <filtered resps for instance 1>]
         """
         return resps
-        
+
+
 @dataclass
 class FilterEnsemble:
     """
     FilterEnsemble creates a pipeline applying multiple filters.
-    Its intended usage is to stack multiple post-processing steps in order. 
-    `task.apply_filters` should use a list of FilterEnsemble classes that it stores, to apply each 
+    Its intended usage is to stack multiple post-processing steps in order.
+    `task.apply_filters` should use a list of FilterEnsemble classes that it stores, to apply each
     pipeline separately.
     """
-    name: str 
+
+    name: str
     filters: List[Filter]
 
     def apply(self, instances: List[Instance]):
 
-        resps = [inst.resps for inst in instances] # operate just on the model responses
+        resps = [
+            inst.resps for inst in instances
+        ]  # operate just on the model responses
         for f in self.filters:
             # apply filters in sequence
             out = f.apply(resps)
-            resps = out # TODO: handle the case where a filter returns multiple "buckets"
-        
+            resps = (
+                out  # TODO: handle the case where a filter returns multiple "buckets"
+            )
+
         # add the end results after filtering to filtered_requests of their respective source instances.
         # has key `self.name`: each FilterEnsemble applied in a given run should use a different name.
         for inst, resp in zip(instances, resps):
             inst.filtered_resps[self.name] = resp
-
-            
-

lm_eval/api/instance.py

 from dataclasses import dataclass, field
 from typing import Literal, Tuple
 
+
 @dataclass
 class Instance:
-    request_type: str = Literal["loglikelihood", "loglikelihood_rolling", "greedy_until"]
+    request_type: str = Literal[
+        "loglikelihood", "loglikelihood_rolling", "greedy_until"
+    ]
     doc: dict = None
     arguments: tuple = None
     idx: int = None
-    metadata: tuple = Tuple[str, int, int] # TODO: better typehints here
+    metadata: tuple = Tuple[str, int, int]  # TODO: better typehints here
     resps: list = field(default_factory=list)
     filtered_resps: dict = field(default_factory=dict)
 
@@ -19,10 +22,12 @@ class Instance:
     def __post_init__(self):
         # unpack metadata field
         self.task_name, self.doc_id, self.repeats = self.metadata
-     
+
     @property
     def args(self):
         """
         Returns (string,) where `string` is the string to calculate loglikelihood over
         """
-        return self.arguments if isinstance(self.arguments, tuple) else (self.arguments,)
+        return (
+            self.arguments if isinstance(self.arguments, tuple) else (self.arguments,)
+        )

lm_eval/api/metrics.py

@@ -10,13 +10,30 @@ import evaluate
 
 
 AGGREGATION_REGISTRY = {}
+
 METRIC_REGISTRY = {
     "acc": None,
     "acc_norm": None,
+    "acc_mutual_info": None,
     "word_perplexity": None,
     "byte_perplexity": None,
 }
 
+HIGHER_IS_BETTER_REGISTRY = {
+    "matthews_corrcoef": True,
+    "f1_score": True,
+    "perplexity": False,
+    "bleu": True,
+    "chrf": True,
+    "ter": False,
+    "acc": True,
+    "acc_norm": True,
+    "acc_mutual_info": True,
+    "word_perplexity": False,
+    "byte_perplexity": False,
+    "bits_per_byte": False,
+}
+
 
 def register_metric(name):
     # TODO: do we want to enforce a certain interface to registered metrics?
@@ -27,7 +44,7 @@ def register_metric(name):
 
         METRIC_REGISTRY[name] = fn
         return fn
-    
+
     return decorate
 
 
@@ -37,12 +54,14 @@ def get_metric(name):
         return METRIC_REGISTRY[name]
     except KeyError:
         # TODO: change this print to logging?
-        print(f"Could not find registered metric '{name}' in lm-eval, \
-searching in HF Evaluate library...")
+        print(
+            f"Could not find registered metric '{name}' in lm-eval, \
+searching in HF Evaluate library..."
+        )
         try:
             metric_object = evaluate.load(name)
             return metric_object.compute
-        except:
+        except Exception:
             raise Warning(
                 "{} not found in the evaluate library!".format(name),
                 "Please check https://huggingface.co/evaluate-metric",
@@ -58,7 +77,7 @@ def register_aggregation(name):
 
         AGGREGATION_REGISTRY[name] = fn
         return fn
-    
+
     return decorate
 
 

lm_eval/api/model.py

@@ -6,14 +6,15 @@ from lm_eval import utils
 
 MODEL_REGISTRY = {}
 
+
 def register_model(*names):
     # either pass a list or a single alias.
     # function receives them as a tuple of strings
-    
+
     def decorate(cls):
-        for name in names: 
-            assert (
-                issubclass(cls, LM)
+        for name in names:
+            assert issubclass(
+                cls, LM
             ), f"Model '{name}' ({cls.__name__}) must extend LM class"
 
             assert (
@@ -22,7 +23,7 @@ def register_model(*names):
 
             MODEL_REGISTRY[name] = cls
         return cls
-    
+
     return decorate
 
 

lm_eval/api/register.py

+import os
+
+task_registry = {}
+group_registry = {}
+task2func_index = {}
+func2task_index = {}
+
+
+def register_task(name):
+    def wrapper(func):
+
+        task_registry[name] = func
+        func2task_index[func.__name__] = name
+        task2func_index[name] = func.__name__
+        return func
+
+    return wrapper
+
+
+def register_group(name):
+    def wrapper(func):
+
+        func_name = func2task_index[func.__name__]
+
+        if name in group_registry:
+            group_registry[name].append(func_name)
+        else:
+            group_registry[name] = [func_name]
+        return func
+
+    return wrapper

lm_eval/api/samplers.py

-
-
 class Sampler:
-
     def __init__(self, docs, task, fewshot_indices=None, rnd=None):
 
         self.rnd = rnd
@@ -12,15 +9,18 @@ class Sampler:
 
         self.delimiter = self.config.delimiter
 
-        self.docs = docs # HF dataset split, provided by task._fewshot_docs()
-        if fewshot_indices: # subset few-shot docs from 
+        self.docs = docs  # HF dataset split, provided by task._fewshot_docs()
+        if fewshot_indices:  # subset few-shot docs from
             self.docs = self.docs.select(fewshot_indices)
 
-
     def get_context(self, doc, num_fewshot):
 
-        # draw an extra fewshot sample if using same split as evaluting on
-        n_samples = num_fewshot + 1 if self.config.fewshot_split == self.config.test_split else num_fewshot 
+        # draw an extra fewshot sample if using same split as evaluating on
+        n_samples = (
+            num_fewshot + 1
+            if self.config.fewshot_split == self.config.test_split
+            else num_fewshot
+        )
 
         # draw `n_samples` docs from fewshot_docs
         fewshotex = self.sample(n_samples)
@@ -28,16 +28,16 @@ class Sampler:
         # get rid of the doc that's the one we're evaluating, if it's in the fewshot
         # TODO: should we just stop people from using fewshot from same split as evaluating?
         selected_docs = [x for x in fewshotex if x != doc][:num_fewshot]
-        
+
         labeled_examples = (
-                self.delimiter.join(
-                    [
-                        self.task.doc_to_text(doc) + self.task.doc_to_target(doc)
-                        for doc in selected_docs
-                    ]
-                )
-                + self.delimiter
+            self.delimiter.join(
+                [
+                    self.task.doc_to_text(doc) + self.task.doc_to_target(doc)
+                    for doc in selected_docs
+                ]
             )
+            + self.delimiter
+        )
 
         # only returns the fewshot context! Does not append the document, do this outside the object
         return labeled_examples
@@ -51,25 +51,22 @@ class Sampler:
 
 
 class BalancedSampler(Sampler):
-
     def sample(self, n):
         """
-        TODO: this should return approximately class-balanced samples from our fewshot examples. 
+        TODO: this should return approximately class-balanced samples from our fewshot examples.
         TODO: what order should they be in? maybe random?
         """
 
         pass
 
-class ManualSampler(Sampler):
 
+class ManualSampler(Sampler):
     def sample(self, n):
-        """
-
-        """
-        pass 
+        """ """
+        pass
 
 
-# TODO: how should we do design here? might be better to have a single sampler and pass more kwargs at init. 
+# TODO: how should we do design here? might be better to have a single sampler and pass more kwargs at init.
 # Depends what's easier for new user to add own functionality on top of
 
 # types of sampler:

lm_eval/evaluator.py

-import collections
+import random
 import itertools
+import collections
+
+import torch
+
 import numpy as np
-import random
+
+import lm_eval.api
 import lm_eval.api.metrics
-import lm_eval.models
+
 import lm_eval.tasks
-import lm_eval.api
-from lm_eval.utils import positional_deprecated, run_task_tests, make_table, create_iterator
-import torch 
+import lm_eval.models
+
+from lm_eval.utils import (
+    positional_deprecated,
+    run_task_tests,
+    make_table,
+    create_iterator,
+    get_git_commit_hash,
+)
+
+from lm_eval.logger import eval_logger
+
 
 @positional_deprecated
 def simple_evaluate(
@@ -65,7 +79,7 @@ def simple_evaluate(
         assert isinstance(model, lm_eval.api.model.LM)
         lm = model
 
-    task_dict = lm_eval.api.task.get_task_dict(tasks, num_fewshot=num_fewshot)
+    task_dict = lm_eval.tasks.get_task_dict(tasks, num_fewshot=num_fewshot)
 
     if check_integrity:
         run_task_tests(task_list=tasks)
@@ -73,7 +87,6 @@ def simple_evaluate(
     results = evaluate(
         lm=lm,
         task_dict=task_dict,
-        num_fewshot=num_fewshot,
         limit=limit,
         bootstrap_iters=bootstrap_iters,
         decontamination_ngrams_path=decontamination_ngrams_path,
@@ -91,13 +104,12 @@ def simple_evaluate(
             "limit": limit,
             "bootstrap_iters": bootstrap_iters,
         }
-
+        results["git_hash"] = get_git_commit_hash()
         return results
     else:
         return None
 
 
-
 decontaminate_suffix = "_decontaminate"
 
 
@@ -105,7 +117,6 @@ decontaminate_suffix = "_decontaminate"
 def evaluate(
     lm,
     task_dict,
-    num_fewshot=0,
     limit=None,
     bootstrap_iters=100000,
     decontamination_ngrams_path=None,
@@ -126,47 +137,54 @@ def evaluate(
         Dictionary of results
     """
 
-    decontaminate = decontamination_ngrams_path is not None
+    # decontaminate = decontamination_ngrams_path is not None
 
     results = collections.defaultdict(dict)
     versions = collections.defaultdict(dict)
 
     requests = collections.defaultdict(list)
-    requests_origin = collections.defaultdict(list)
+    # requests_origin = collections.defaultdict(list)
 
-    docs = {}
+    # docs = {}
 
     # get lists of each type of request
     for task_name, task in task_dict.items():
         versions[task_name] = task.VERSION
-    
+
         # deterministically shuffle docs and chop off the first `limit` because sometimes docs are in some kind of order
         # task_docs = list(task_doc_func())
         # rnd = random.Random()
         # rnd.seed(42)
         # rnd.shuffle(task_docs)
 
-        task.build_all_requests(limit=limit, rank = lm.rank, world_size = lm.world_size)
+        task.build_all_requests(limit=limit, rank=lm.rank, world_size=lm.world_size)
+
         # aggregate Instances by LM method requested to get output.
-        reqtype = "loglikelihood" if task.OUTPUT_TYPE == "multiple_choice" else task.OUTPUT_TYPE #TODO: this is hacky, fix in task.py
-        requests[reqtype].extend(task.instances) 
+        reqtype = (
+            "loglikelihood"
+            if task.OUTPUT_TYPE == "multiple_choice"
+            else task.OUTPUT_TYPE
+        )  # TODO: this is hacky, fix in task.py
+        requests[reqtype].extend(task.instances)
 
         if lm.world_size > 1:
-            instances_rnk = torch.tensor(len(task._instances), device = lm.device)
-            gathered_item = lm.accelerator.gather(instances_rnk).cpu().detach().numpy().tolist()
+            instances_rnk = torch.tensor(len(task._instances), device=lm.device)
+            gathered_item = (
+                lm.accelerator.gather(instances_rnk).cpu().detach().numpy().tolist()
+            )
 
             # compute number of pseudobatches to pad with (FSDP/DDP require even batches among ranks)
             numpad = max(gathered_item) - gathered_item[lm.rank]
-    
+
     ### Run LM on inputs, get all outputs ###
     # execute each type of request
     for reqtype, reqs in requests.items():
-        print("Running", reqtype, "requests")
+        eval_logger.info("Running {} requests".format(reqtype))
         # create `K` copies of each request `req` based off `K = req.repeats`
         cloned_reqs = []
         for req in reqs:
             cloned_reqs.extend([req] * req.repeats)
-        
+
         if (lm.world_size > 1) and (numpad > 0):
             for _ in range(numpad):
                 cloned_reqs.extend([req] * req.repeats)
@@ -186,9 +204,8 @@ def evaluate(
     for task_name, task in task_dict.items():
         task.apply_filters()
 
-
     ### Collect values of metrics on all datapoints ###
-    # TODO: make metric configurable, add metric registry 
+    # TODO: make metric configurable, add metric registry
     vals = collections.defaultdict(list)
 
     # unpack results and sort back in order and return control to Task
@@ -196,55 +213,70 @@ def evaluate(
         # calculate values for each filter setup (TODO: make getting list of keys cleaner)
         # TODO: make it possible to use a different metric per key
         for key in task.instances[0].filtered_resps.keys():
-            doc_iterator = itertools.islice(enumerate(task.test_docs()), lm.rank, limit, lm.world_size) if task.has_test_docs() else itertools.islice(enumerate(task.validation_docs()), lm.rank, limit, lm.world_size)
+            doc_iterator = (
+                itertools.islice(
+                    enumerate(task.test_docs()), lm.rank, limit, lm.world_size
+                )
+                if task.has_test_docs()
+                else itertools.islice(
+                    enumerate(task.validation_docs()), lm.rank, limit, lm.world_size
+                )
+            )
             for doc_id, doc in doc_iterator:
                 # subset instances to only this document id ; sort by idx
                 requests = list(filter(lambda x: x.doc_id == doc_id, task.instances))
                 requests.sort(key=lambda x: x.idx)
-                metrics = task.process_results(doc, [req.filtered_resps[key] for req in requests])
+                metrics = task.process_results(
+                    doc, [req.filtered_resps[key] for req in requests]
+                )
                 for metric, value in metrics.items():
                     vals[(task_name, key, metric)].append(value)
-    
+
     if lm.world_size > 1:
-        # if multigpu, then gather data across all ranks    
+        # if multigpu, then gather data across all ranks
         vals_torch = collections.defaultdict(list)
         for (task_name, key, metric), items in vals.items():
-            
-            numitem = 0 
+
+            numitem = 0
             if type(items[0]) == tuple:
-                numitem = len(items[0]) 
-    
+                numitem = len(items[0])
+
             # distributed gather requires all ranks to have same dimensions
             # so we pad out with float32 min value
             pad_value = torch.finfo(torch.float32).min
-            metrics_tensor = torch.tensor(items, device = lm.device)
-            
-            original_dtype = metrics_tensor.dtype # store original dtype 
-            torch_device_tensor = lm.accelerator.pad_across_processes(metrics_tensor.to(torch.float32), pad_index = pad_value)
+            metrics_tensor = torch.tensor(items, device=lm.device)
+
+            original_dtype = metrics_tensor.dtype  # store original dtype
+            torch_device_tensor = lm.accelerator.pad_across_processes(
+                metrics_tensor.to(torch.float32), pad_index=pad_value
+            )
             gathered_item = lm.accelerator.gather(torch_device_tensor)
-    
+
             if numitem > 0:
-                gathered_filtered = gathered_item[gathered_item[:,0] != pad_value]
+                gathered_filtered = gathered_item[gathered_item[:, 0] != pad_value]
             else:
                 gathered_filtered = gathered_item[gathered_item != pad_value]
-                
-            gathered_item = gathered_filtered.to(original_dtype).cpu().detach().numpy().tolist()
+
+            gathered_item = (
+                gathered_filtered.to(original_dtype).cpu().detach().numpy().tolist()
+            )
             # reconvert if we were passed a tuple of values
             if numitem > 0:
                 gathered_item = [tuple(g) for g in gathered_item]
-    
+
             if lm.rank == 0:
                 vals_torch[(task_name, key, metric)] = gathered_item
-    
-        vals = vals_torch
 
+        vals = vals_torch
 
     if lm.rank == 0:
         ### Aggregate results over all datapoints ###
         # aggregate results ; run bootstrap CIs
         for (task_name, key, metric), items in vals.items():
             task = task_dict[task_name]
-            results[task_name][metric + " - filter=" + key] = task.aggregation()[metric](items)
+            results[task_name][metric + " - filter=" + key] = task.aggregation()[
+                metric
+            ](items)
 
             # hotfix: bleu, chrf, ter seem to be really expensive to bootstrap
             # so we run them less iterations. still looking for a cleaner way to do this
@@ -257,7 +289,9 @@ def evaluate(
             )
 
             if stderr is not None:
-                results[task_name][metric + " - filter=" + key + "_stderr"] = stderr(items)
+                results[task_name][metric + " - filter=" + key + "_stderr"] = stderr(
+                    items
+                )
 
         return {"results": dict(results), "versions": dict(versions)}
 

lm_eval/filters/__init__.py

@@ -6,10 +6,10 @@ from . import extraction
 FILTER_REGISTRY = {
     "take_first": selection.TakeFirstFilter,
     "regex": extraction.RegexFilter,
-    # TODO: implement this filter. either it should take in an arbitrary "scoring"/reward function 
-    # that takes an input and returns a scalar and then should select the max reward, 
+    # TODO: implement this filter. either it should take in an arbitrary "scoring"/reward function
+    # that takes an input and returns a scalar and then should select the max reward,
     # or should implement different filters for different ways of handling a reward model's inference.
-    #"arg_max": selection.ArgMaxFilter, 
+    # "arg_max": selection.ArgMaxFilter,
 }
 
 
@@ -17,16 +17,19 @@ def get_filter(filter_name):
     return FILTER_REGISTRY[filter_name]
 
 
-def build_filter_ensemble(name, components):
+def build_filter_ensemble(filter_name, components):
     """
     Create a filtering pipeline.
     """
+
     filters = []
-    for step in components:
+    for (function, kwargs) in components:
+        if kwargs is None:
+            f = get_filter(function)()
+        else:
             # create a filter given its name in the registry
-            f = get_filter(step)() # TODO: pass kwargs to filters properly
-
-            # add the filter as a pipeline step
-            filters.append(f)
+            f = get_filter(function)(**kwargs)  # TODO: pass kwargs to filters properly
+        # add the filter as a pipeline step
+        filters.append(f)
 
-    return FilterEnsemble(name=name, filters=filters)
+    return FilterEnsemble(name=filter_name, filters=filters)

lm_eval/filters/decontamination.py

@@ -4,7 +4,7 @@ from lm_eval.api.filter import Filter
 class DecontaminationFilter(Filter):
 
     """
-    A filter which evaluates 
+    A filter which evaluates
     """
 
     name = "track_decontamination"
@@ -12,7 +12,7 @@ class DecontaminationFilter(Filter):
     def __init__(self, path):
         """
 
-        TODO: make sure only ever run one time on the train set (should this be cached as a class var? keyed by value for "path"). 
+        TODO: make sure only ever run one time on the train set (should this be cached as a class var? keyed by value for "path").
         should further cache result on a given (task_name, doc_id)
         """
         self._decontam_results = None
@@ -21,4 +21,4 @@ class DecontaminationFilter(Filter):
         """
         Return {"no_contamination", "only_contamination"} keys for the 2 different subsets
         """
-        pass
\ No newline at end of file
+        pass

lm_eval/filters/extraction.py

@@ -4,24 +4,20 @@ from lm_eval.api.filter import Filter
 
 
 class RegexFilter(Filter):
-    """
+    """ """
 
-
-    """
-
-    def __init__(self, regex=r"#### (\-?[0-9\.\,]+)", fallback="[invalid]"):
+    def __init__(self, regex_pattern=r"#### (\-?[0-9\.\,]+)", fallback="[invalid]"):
         """
         pass a string `regex` to run `re.compile(r"regex")` on.
         `fallback` defines the output returned if no matches for the regex are located.
         """
-        self.regex_pattern = regex
-        self.regex = re.compile(regex)
-
+        self.regex_pattern = regex_pattern
+        self.regex = re.compile(regex_pattern)
         self.fallback = fallback
 
     def apply(self, resps):
         # here, we assume we have a list, in which each element is
-        # a list of model responses for some particular input/target pair. 
+        # a list of model responses for some particular input/target pair.
         # so we process each of these (same input/target response sets)
         # independently (and keep them a list.)
         def filter_set(inst):
@@ -30,7 +26,7 @@ class RegexFilter(Filter):
                 match = self.regex.search(resp)
                 if match:
                     match = match.group(1).strip()
-                    match_str.replace(",", "")
+                    match.replace(",", "")
                     # TODO: should we assume any other filtering is performed?
                 else:
                     match = self.fallback

lm_eval/filters/selection.py

 from lm_eval.api.filter import Filter
 
-class TakeFirstFilter:
 
+class TakeFirstFilter:
     def __init__(self):
         """
         Can define custom behavior here, if an individual instantiation of a Filter class should have state.
@@ -11,4 +11,4 @@ class TakeFirstFilter:
         """
         Assuming each entry of `resps` is a list of model responses, we discard all but the first response.
         """
-        return map(lambda r: r[0], resps)
\ No newline at end of file
+        return map(lambda r: r[0], resps)

lm_eval/logger.py

+import logging
+
+logging.basicConfig(
+    format="%(asctime)s,%(msecs)03d %(levelname)-8s [%(filename)s:%(lineno)d] %(message)s",
+    datefmt="%Y-%m-%d:%H:%M:%S",
+    level=logging.INFO,
+)
+eval_logger = logging.getLogger("lm-eval")

lm_eval/models/__init__.py

-from lm_eval.api.model import LM, MODEL_REGISTRY
-
 from . import gpt2
 from . import gpt3
 from . import textsynth

lm_eval/models/gpt2.py

@@ -6,11 +6,13 @@ from tqdm import tqdm
 import torch.nn.functional as F
 
 from lm_eval import utils
+from lm_eval.logger import eval_logger
 from lm_eval.api.model import LM, register_model
 
 from accelerate import Accelerator
 from itertools import islice
 
+
 @register_model("hf-causal", "gpt2")
 class HFLM(LM):
     def __init__(
@@ -28,10 +30,10 @@ class HFLM(LM):
         assert isinstance(device, str)
         assert isinstance(pretrained, str)
         assert isinstance(batch_size, int)
-        
+
         gpus = torch.cuda.device_count()
         if gpus <= 1:
-            if device:  
+            if device:
                 if device not in ["cuda", "cpu"]:
                     device = int(device)
                 self._device = torch.device(device)
@@ -48,7 +50,7 @@ class HFLM(LM):
             self._world_size = 1
 
         else:
-            self._device = 'cpu'
+            self._device = "cpu"
 
         # TODO: update this to be less of a hack once subfolder is fixed in HF
         revision = revision + ("/" + subfolder if subfolder is not None else "")
@@ -72,14 +74,16 @@ class HFLM(LM):
         if gpus > 1:
             accelerator = Accelerator(device_placement=False)
             if gpus > accelerator.num_processes:
-                warning = ("WARNING: The number of total GPUs does not match the number of spawned processes. " 
-                      "If you would like to use data parallelism, please launch the script "
-                      "with 'accelerate launch *script*'. " 
-                        "Current run will proceed with single device.")
+                warning = (
+                    "WARNING: The number of total GPUs does not match the number of spawned processes. "
+                    "If you would like to use data parallelism, please launch the script "
+                    "with 'accelerate launch *script*'. "
+                    "Current run will proceed with single device."
+                )
                 print(warning)
                 self._rank = 0
                 self._world_size = 1
-            
+
             else:
                 self.gpt2 = accelerator.prepare(self.gpt2)
                 self._device = torch.device(f"cuda:{accelerator.local_process_index}")
@@ -90,7 +94,6 @@ class HFLM(LM):
 
                 self._rank = self.accelerator.local_process_index
                 self._world_size = self.accelerator.num_processes
-            
 
     @property
     def eot_token_id(self):
@@ -116,7 +119,7 @@ class HFLM(LM):
     @property
     def device(self):
         return self._device
-    
+
     @property
     def rank(self):
         return self._rank
@@ -144,7 +147,11 @@ class HFLM(LM):
 
     def _model_generate(self, context, max_length, eos_token_id):
         return self.gpt2.generate(
-            context, max_length=max_length, pad_token_id=eos_token_id, eos_token_id=eos_token_id, do_sample=False
+            context,
+            max_length=max_length,
+            pad_token_id=eos_token_id,
+            eos_token_id=eos_token_id,
+            do_sample=False,
         )
 
     def loglikelihood(self, requests):
@@ -168,11 +175,11 @@ class HFLM(LM):
 
         extra_pad = []
         numpad_batches = 0
-        
+
         if self.world_size > 1:
             cumulative_batches = 0  # balance token batches among iterators
-            # compute cumlative batches seen per host
-            for (string,) in tqdm([req.args for req in requests],disable=True):
+            # compute cumulative batches seen per host
+            for (string,) in tqdm([req.args for req in requests], disable=True):
                 rolling_token_windows = list(
                     map(
                         utils.make_disjoint_window,
@@ -188,30 +195,42 @@ class HFLM(LM):
                 rolling_token_windows = [(None,) + x for x in rolling_token_windows]
                 cumulative_batches += len(rolling_token_windows)
 
-            cumul_batches_ranks = torch.tensor(cumulative_batches, device = self.device)
-            gathered_item = self.accelerator.gather(cumul_batches_ranks).cpu().detach().numpy().tolist()
+            cumul_batches_ranks = torch.tensor(cumulative_batches, device=self.device)
+            gathered_item = (
+                self.accelerator.gather(cumul_batches_ranks)
+                .cpu()
+                .detach()
+                .numpy()
+                .tolist()
+            )
 
             # compute number of pseudobatches to pad with (FSDP/DDP require even batches among ranks)
             numpad_batches = max(gathered_item) - gathered_item[self.rank]
 
-            # pad iterators with a pseudodocument 
-            extra_pad = [('pad',)] if max(gathered_item) - min(gathered_item) > 0 else []
+            # pad iterators with a pseudodocument
+            extra_pad = (
+                [("pad",)] if max(gathered_item) - min(gathered_item) > 0 else []
+            )
 
         loglikelihoods = []
-        for (string,) in tqdm(extra_pad + [req.args for req in requests],disable=(self.rank != 0)):
+        for (string,) in tqdm(
+            extra_pad + [req.args for req in requests], disable=(self.rank != 0)
+        ):
             if numpad_batches > 0:
                 rolling_token_windows = list(
                     map(
                         utils.make_disjoint_window,
                         utils.get_rolling_token_windows(
-                            token_list=[self.eot_token_id]*self.max_length*numpad_batches,
+                            token_list=[self.eot_token_id]
+                            * self.max_length
+                            * numpad_batches,
                             prefix_token=self.eot_token_id,
                             max_seq_len=self.max_length,
                             context_len=1,
                         ),
                     )
                 )
-                
+
             else:
                 rolling_token_windows = list(
                     map(
@@ -233,8 +252,8 @@ class HFLM(LM):
                 rolling_token_windows, disable_tqdm=True
             )
 
-            if (numpad_batches > 0) or (string == 'pad'):
-                numpad_batches = 0 
+            if (numpad_batches > 0) or (string == "pad"):
+                numpad_batches = 0
 
             else:
                 # discard is_greedy
@@ -242,7 +261,6 @@ class HFLM(LM):
 
                 string_nll = sum(string_nll)
                 loglikelihoods.append(string_nll)
-        
 
         return loglikelihoods
 
@@ -264,7 +282,8 @@ class HFLM(LM):
         # TODO: automatic (variable) batch size detection for vectorization
         re_ord = utils.Reorderer(requests, _collate)
         for chunk in utils.chunks(
-            tqdm(re_ord.get_reordered(), disable=(disable_tqdm or (self.rank != 0))), self.batch_size
+            tqdm(re_ord.get_reordered(), disable=(disable_tqdm or (self.rank != 0))),
+            self.batch_size,
         ):
             inps = []
             cont_toks_list = []

lm_eval/models/gpt3.py

 import os
-import numpy as np
+import time
 import transformers
-from lm_eval.api.model import LM, register_model
-from lm_eval import utils
+
+import numpy as np
+
 from tqdm import tqdm
-import time
+from lm_eval import utils
+from lm_eval.api.model import LM, register_model
 
 
 def get_result(response, ctxlen):