fix issue such that both l1_l2 and l2_l1 work

ted_talks.py

@@ -23,8 +23,7 @@ class TedTalksConfig(datasets.BuilderConfig):
     """BuilderConfig for TED talk dataset."""
 
     def __init__(self, language_pair=(None, None), **kwargs):
-        # sort such that az_ar is same as ar_az
-        self.language_pair = sorted(language_pair)
+        self.language_pair = language_pair
         self.source, self.target = self.language_pair[0], self.language_pair[1]
 
         name = f"{self.source}_{self.target}"
@@ -35,12 +34,12 @@ class TedTalksConfig(datasets.BuilderConfig):
 class TedTalks(datasets.GeneratorBasedBuilder):
     """TED talk data from http://phontron.com/data/ted_talks.tar.gz."""
 
-    unique_pairs = sorted(set([
-        "_".join(sorted([l1, l2]))
+    unique_pairs = [
+        "_".join([l1, l2])
         for l1 in _LANGUAGES
         for l2 in _LANGUAGES
         if l1 != l2
-    ]))
+    ]
 
     BUILDER_CONFIGS = [
         TedTalksConfig(
@@ -150,4 +149,4 @@ class TedTalks(datasets.GeneratorBasedBuilder):
             source_sentences, target_sentences = zip(*source_target_pairs)
 
             for idx, (l1, l2) in enumerate(zip(source_sentences, target_sentences)):
-                yield idx, {"translation": {self.config.source: l1, self.config.target: l2}}
+                yield idx, {"translation": {self.config.source: l1, self.config.target: l2}}