Merge pull request #103 from EleutherAI/piqa

Implement PiQA

lm_eval/base.py

@@ -59,7 +59,6 @@ class LM(abc.ABC):
 
 
 class Dataset(abc.ABC):
-    @abc.abstractmethod
     def __init__(self):
         self.download()
         self._traindocs = None

lm_eval/tasks/__init__.py

@@ -14,6 +14,7 @@ from . import naturalqs
 from . import sat
 from . import arithmetic
 from . import lambada
+from . import piqa
 
 TASK_REGISTRY = {
     # GLUE
@@ -39,6 +40,7 @@ TASK_REGISTRY = {
     # Order by benchmark/genre?
 
     "lambada": lambada.LAMBADA,
+    "piqa": piqa.PiQA,
 
     # "arc_easy": arc.ARCEasy, # not implemented yet
     # "arc_challenge": arc.ARCChallenge, # not implemented yet

lm_eval/tasks/piqa.py

 import json
 import random
-from lm_eval.base import Dataset
+from lm_eval.base import Dataset, rf, mean
 from ..utils import sh
+import os
 
 class PiQA(Dataset):
-    def __init__(self):
-        self.download()
     def download(self):
-        #pass
-        #TODO: don't download if files already there
+        if not os.path.exists('data/piqa'):
+            #TODO: use best_download
             sh("""
             mkdir -p data/piqa
             wget https://yonatanbisk.com/piqa/data/train.jsonl -O data/piqa/piqa-train.jsonl
@@ -25,11 +24,11 @@ class PiQA(Dataset):
         return True
 
     def has_test_docs(self):
-        return True
+        return False
 
     def load_docs(self, textfilename, labelfilename):
         if labelfilename != None:
-            return zip([json.loads(entry) for entry in list(open(textfilename,'r'))],list(open(labelfilename, 'r')))
+            return zip([json.loads(entry) for entry in list(open(textfilename,'r'))],list(map(lambda x: x.strip(), open(labelfilename, 'r'))))
         else:
             return [json.loads(entry) for entry in list(open(textfilename,'r'))]
     
@@ -39,62 +38,40 @@ class PiQA(Dataset):
     def validation_docs(self):
         return self.load_docs('data/piqa/piqa-valid.jsonl', 'data/piqa/piqa-valid-labels.lst')
 
-    def test_docs(self):
-        return self.load_docs('data/piqa/piqa-test.jsonl', None)
+    #def test_docs(self):
+    #    return self.load_docs('data/piqa/piqa-test.jsonl', None)
     
     def fewshot_description(self):
         # TODO: figure out fewshot description
         return ""
     
     def doc_to_text(self, doc):
-        #TODO: check if oa uses newline
-        return  doc['goal'] + ' '
+        return doc[0]['goal']
 
     def doc_to_target(self, doc):
-        rightanswer = int(doc[1][0]) + 1
-        return ''.join([doc[0]['goal'],' ',doc[0]['sol'+str(rightanswer)]])
+        #TODO: check if oa uses newline
+        rightanswer = int(doc[1]) + 1
+        return '\n' + ''.join([doc[0]['goal'],' ',doc[0]['sol'+str(rightanswer)]])
 
     def construct_requests(self, doc, ctx):
-        """ 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`. 
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        ll_1, _ = rf.loglikelihood(ctx, doc[0]['sol1'])
+        ll_2, _ = rf.loglikelihood(ctx, doc[0]['sol2'])
+
+        return ll_1, ll_2
     
     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.
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        ll_1, ll_2 = results
+
+        return {
+            'acc': (ll_1 > ll_2) == (int(doc[1]) == 0)
+        }
 
     def aggregation(self):
-        """
-        :returns: {str: [float] -> float}
-            A dictionary where keys are the names of submetrics and values are 
-            functions that aggregate a list of metrics
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
+        return {
+            'acc': mean
+        }
 
     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
+        return {
+            'acc': True
+        }
\ No newline at end of file
-        """
-        # TODO: implement evaluation.
-        raise NotImplementedError('Evaluation not implemented')
\ No newline at end of file