Fixed XNLI prompts entailment direction (#52)

* fixed subject-verb agreement in XNLI prompt

* fixed NLI direction: P entails H, not vice versa

We could maybe either prompt "does S1 contradict S2" or “does S2 contradict S1”. But for entailment we can only prompt “does S1 entail S2” (does the premise entail the hypothesis) not the other way around.

* standardized line breaks and oxford commas

To exactly match Figure G7 in the GPT-3 paper.

templates/xnli/en/templates.yaml

@@ -3,38 +3,38 @@ subset: en
 templates:
   4e122d26-7e79-49c0-961b-cf8ee134759e: !Template
     id: 4e122d26-7e79-49c0-961b-cf8ee134759e
-    jinja: "Sentence 1: {{premise}}\nSentence 2: {{hypothesis}}\n\nDoes Sentence 2\
-      \ contradicts Sentence 1? Yes, No or Neutral? |||\n{% if label == 0 %} \nNo\n\
-      {% elif label == 1 %}\nNeutral\n{% else %}\nYes\n{% endif %}"
+    jinja: "Sentence 1: {{premise}}\nSentence 2: {{hypothesis}}\nQuestion: Does Sentence\
+      \ 1 contradict Sentence 2? Yes, No, or Neutral? |||\n{% if label == 0 %} \n\
+      No\n{% elif label == 1 %}\nNeutral\n{% else %}\nYes\n{% endif %}"
     name: Concatenation contraposition
     reference: Concatenation contraposition
   c62a3048-018e-4d93-bc46-645f3f763ee6: !Template
     id: c62a3048-018e-4d93-bc46-645f3f763ee6
-    jinja: "Sentence 1: {{premise}}\n\nSentence 2: {{hypothesis}}\n\nDoes Sentence\
-      \ 2 contradicts Sentence 1? Yes or No? |||\n{% if label == 2 %} \nYes\n{% else\
+    jinja: "Sentence 1: {{premise}}\nSentence 2: {{hypothesis}}\nQuestion: Does Sentence\
+      \ 1 contradict Sentence 2? Yes or No? |||\n{% if label == 2 %} \nYes\n{% else\
       \ %}\nNo\n{% endif %}"
-    name: Label binarization  contraposition
+    name: Label binarization contraposition
     reference: Inspired from https://arxiv.org/pdf/1902.01007.pdf Section 4 - Implementation
       and evaluation
   d9e13133-267e-46c4-afad-c2379dcc5272: !Template
     id: d9e13133-267e-46c4-afad-c2379dcc5272
-    jinja: "{{premise}}\nQuestion: {{hypothesis}}\nTrue, False or Neither? ||| \n\
+    jinja: "{{premise}}\nQuestion: {{hypothesis}} True, False, or Neither? ||| \n\
       {% if label == 0 %} \nTrue\n{% elif label == 1 %}\nNeither\n{% else %}\nFalse\n\
       {% endif %}"
     name: ANLI GPT3
     reference: ANLI prompt format from Table G7 in the GPT3 paper
   dd4276e6-aebd-44a3-b3cf-baf8a4c237f0: !Template
     id: dd4276e6-aebd-44a3-b3cf-baf8a4c237f0
-    jinja: "Sentence 1: {{premise}}\n\nSentence 2: {{hypothesis}}\n\nDoes Sentence\
-      \ 2 entails Sentence 1? Yes or No? |||\n{% if label == 0 %} \nYes\n{% else %}\n\
+    jinja: "Sentence 1: {{premise}}\nSentence 2: {{hypothesis}}\nQuestion: Does Sentence\
+      \ 1 entail Sentence 2? Yes or No? |||\n{% if label == 0 %} \nYes\n{% else %}\n\
       No\n{% endif %}"
     name: Label binarization
     reference: Grouping "neutral" and "contradiction" as a single label following
       https://arxiv.org/pdf/1902.01007.pdf Section 4 - Implementation and evaluation
   e174f56a-b0af-4937-b6ae-1897cac26eba: !Template
     id: e174f56a-b0af-4937-b6ae-1897cac26eba
-    jinja: "Sentence 1: {{premise}}\nSentence 2: {{hypothesis}}\n\nDoes Sentence 2\
-      \ entails Sentence 1? Yes, No or Neutral? |||\n{% if label == 0 %} \nYes\n{%\
-      \ elif label == 1 %}\nNeutral\n{% else %}\nNo\n{% endif %}"
+    jinja: "Sentence 1: {{premise}}\nSentence 2: {{hypothesis}}\nQuestion: Does Sentence\
+      \ 1 entail Sentence 2? Yes, No, or Neutral? |||\n{% if label == 0 %} \nYes\n\
+      {% elif label == 1 %}\nNeutral\n{% else %}\nNo\n{% endif %}"
     name: Concatenation
     reference: Concatenation of premise and hypothesis