update coqa to be consistent with gpt3 paper

lm_eval/tasks/coqa.py

@@ -5,8 +5,10 @@ from ..utils import sh
 
 
 class CoQA(Dataset):
-
+    def __init__(self):
+        self.download()
     def download(self):
+        #TODO: don't download if files already there
         sh("""
             mkdir -p data/coqa 
             wget http://downloads.cs.stanford.edu/nlp/data/coqa/coqa-train-v1.0.json -O data/coqa/coqa-train-v1.0.json
@@ -23,42 +25,25 @@ class CoQA(Dataset):
         return False
 
     def training_docs(self):
-        myjson = json.load(open('data/coqa/coqa-train-v1.0.json'))['data']
-        return self.load_doc(myjson)
+        return json.load(open('data/coqa/coqa-train-v1.0.json'))['data']
 
     def validation_docs(self):
         pass
 
     def test_docs(self):
-        myjson = json.load(open('data/coqa/coqa-dev-v1.0.json'))['data']    
-        return self.load_doc(myjson)
-    
-    def fewshot_examples(self, k):
-        traindocs = list(self.training_docs())
-        random.seed(123)
-        random.shuffle(traindocs)
-
-        return traindocs[:k]
+        return  json.load(open('data/coqa/coqa-dev-v1.0.json'))['data']     
     
     def fewshot_description(self):
         pass
-
-    def load_doc(self, myjson):
-        docs = []
-        for item in myjson:
-            new_instance = [item['story']]
-            qa_pairs = zip(item['questions'], item['answers'])
-            for pair in qa_pairs:
-                new_instance.append('\n')
-                new_instance.append(''.join(['Q: ',pair[0]['input_text']]))
-                new_instance.append(''.join(['A: ',pair[1]['input_text']]))
-            docs.append(new_instance)  
-        return docs
     
     def doc_to_text(self, doc, include_target=True):
-        text = '\n<|endoftext|>\n'.join(['\n'.join(instance) for instance in doc])
-        text = text + '\n<|endoftext|>'
-        return text
+        text = [doc['story']]
+        for pair in zip(doc['questions'], doc['answers']):
+            text.append('\n\n')
+            text.append(''.join(['Q: ',pair[0]['input_text'], '\n\n']))
+            text.append(''.join(['A: ',pair[1]['input_text']]))
+
+        return ''.join(text)
 
     def evaluate(self, docs, lm):
         pass