delete training examples without negatives

create_miracl_japanese_small.py

@@ -12,4 +12,5 @@ if __name__ == '__main__':
     for split in OUTPUT_FILEPATHS:
         with open(OUTPUT_FILEPATHS[split], 'w', encoding='utf-8') as f:
             for data in tqdm(dataset[split]):
-                f.write(json.dumps(data) + '\n')
+                if not split == 'train' or len(data['negative_passages']) > 0:
+                    f.write(json.dumps(data) + '\n')

topics.miracl-v1.0-ja-train.jsonl.gz

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:acf7646d6a6d6de749f2cb0ce4ba1fda5e7a2085a5abc958225572fca5d8ae73
-size 12545689
+oid sha256:a6124e784fd0ffdffcf1689d5090dc6961a8be889791bc04ec438aafcaf58276
+size 12469053