Merge pull request #119 from jon-tow/task-refactor

Refactor `Dataset` naming and `HFTask` properties

lm_eval/base.py

@@ -58,10 +58,10 @@ class LM(abc.ABC):
         return cls()
 
 
-class Dataset(abc.ABC):
+class Task(abc.ABC):
     def __init__(self):
         self.download()
-        self._traindocs = None
+        self._training_docs = None
 
     def download(self):
         """Downloads the task dataset if necessary"""
@@ -84,23 +84,29 @@ class Dataset(abc.ABC):
 
     def training_docs(self):
         """
-
         :return: Iterable[obj]
             A iterable of any object, that doc_to_text can handle
         """
         return []
 
     def validation_docs(self):
+        """
+        :return: Iterable[obj]
+            A iterable of any object, that doc_to_text can handle
+        """
         return []
 
     def test_docs(self):
+        """
+        :return: Iterable[obj]
+            A iterable of any object, that doc_to_text can handle
+        """
         return []
 
     def fewshot_examples(self, k):
-        if self._traindocs is None:
-            self._traindocs = list(self.training_docs())
-
-        return random.sample(self._traindocs, k)
+        if self._training_docs is None:
+            self._training_docs = list(self.training_docs())
+        return random.sample(self._training_docs, k)
 
     @abc.abstractmethod
     def doc_to_text(self, doc):

lm_eval/tasks/arithmetic.py

@@ -2,12 +2,12 @@ import abc
 import json
 import os
 from collections import namedtuple
-from lm_eval.base import Dataset, mean, rf
+from lm_eval.base import Task, mean, rf
 from best_download import download_file
 
 ArithmeticDoc = namedtuple('ArithmeticDoc', ['context', 'completion'])
 
-class Arithmetic(Dataset):
+class Arithmetic(Task):
     directory = 'data/arithmetic/'
 
     def __init__(self):

lm_eval/tasks/common.py

 import datasets
 import numpy as np
-import random
-from ..base import Dataset
+from ..base import Task
 
 
-class HFTask(Dataset):
+class HFTask(Task):
     DATASET_PATH = None
     DATASET_NAME = None
 
     def __init__(self):
+        self.data = None
         super().__init__()
-        self._training_docs = None
 
     def download(self):
         self.data = datasets.load_dataset(path=self.DATASET_PATH, name=self.DATASET_NAME)

lm_eval/tasks/coqa.py

@@ -2,11 +2,11 @@
 
 import json
 import random
-from lm_eval.base import Dataset
+from lm_eval.base import Task
 from ..utils import sh
 
 
-class CoQA(Dataset):
+class CoQA(Task):
     def __init__(self):
         self.download()
     def download(self):

lm_eval/tasks/drop.py

@@ -5,9 +5,9 @@ from sklearn.metrics import f1_score, matthews_corrcoef
 from tqdm import auto as tqdm_lib
 from . common import HFTask, simple_accuracy_metric, yesno
 from pathlib import Path
-from ..base import Dataset
+from ..base import Task
 
-class DROP(Dataset):
+class DROP(Task):
     DATAFOLDER = Path(__file__).parent / "../../data/drop"
     
     def __init__(self):

lm_eval/tasks/lambada.py

-from lm_eval.base import Dataset, rf, mean
+from lm_eval.base import Task, rf, mean
 from lm_eval.utils import sh
 import json
 import math
 from best_download import download_file
 
 
-class LAMBADA(Dataset):
+class LAMBADA(Task):
     def download(self):
         sh("mkdir -p data/lambada")
         download_file(

lm_eval/tasks/naturalqs.py

@@ -30,10 +30,10 @@ class NaturalQs(HFTask):
 
     def fewshot_examples(self, k):
         # Data is too large to fit in memory. We just sample from the first bit.
-        if self._traindocs is None:
-            self._traindocs = list(islice(self.training_docs(), 0, 100000))
+        if self._training_docs is None:
+            self._training_docs = list(islice(self.training_docs(), 0, 100000))
 
-        return random.sample(self._traindocs, k)
+        return random.sample(self._training_docs, k)
 
     def doc_to_text(self, doc):
         return 'Q: ' + doc['question']['text'] + '\n\n' + 'A: '

lm_eval/tasks/piqa.py

 import json
 import random
-from lm_eval.base import Dataset, rf, mean
+from lm_eval.base import Task, rf, mean
 from ..utils import sh
 import os
 
-class PiQA(Dataset):
+class PiQA(Task):
     def download(self):
         if not os.path.exists('data/piqa'):
             #TODO: use best_download

lm_eval/tasks/quac.py

 import json
 import random
 import os
-from lm_eval.base import Dataset
+from lm_eval.base import Task
 from ..utils import sh
 
 
-class QuAC(Dataset):    
+class QuAC(Task):    
     def __init__(self):
         super().__init__()
 

lm_eval/tasks/sat.py

 import json
 import random
 import os
-from lm_eval.base import Dataset, rf, mean
+from lm_eval.base import Task, rf, mean
 from tqdm import auto as tqdm_lib
 from . common import simple_accuracy_metric
 import numpy as np
 from ..utils import sh
 
 
-class SATAnalogies(Dataset):    
+class SATAnalogies(Task):    
     NEEDS_MANUAL_DL = True
     
     def __init__(self):

lm_eval/tasks/storycloze.py

 import json
 import random
-from lm_eval.base import Dataset
+from lm_eval.base import Task
 from ..utils import sh
 import csv
 
-class StoryCloze(Dataset):
+class StoryCloze(Task):
     NEEDS_MANUAL_DL = True
 
     def download(self):

lm_eval/tasks/triviaqa.py

 import os
 import json
 import random
-from lm_eval.base import Dataset, mean, rf
+from lm_eval.base import Task, mean, rf
 from ..utils import sh
 
-class TriviaQA(Dataset):
+class TriviaQA(Task):
     def download(self):
         if not os.path.exists('data/triviaqa'):
             sh("""

lm_eval/tasks/wsc273.py

 import json
 import random
 import os
-from lm_eval.base import Dataset
+from lm_eval.base import Task
 from ..utils import sh
 
 
-class WinogradSchemaChallenge273(Dataset):    
+class WinogradSchemaChallenge273(Task):    
     def __init__(self):
         super().__init__()