Duplicate negatives appearing within single rows of triplet-5
Thank you very much for releasing this very useful dataset.
While working with the triplet-5 (train) split, I checked the five negative columns (negative_1–negative_5) across all 96,658 rows and noticed that duplicates appear within a single row. In particular, some rows contain all negatives that are exactly the same text. The counts I observed are:
- All five negatives identical (5): 400 rows (≈0.41%)
- Four identical negatives (4,1): 774 rows (≈0.80%)
- Three identical negatives (3,2 or 3,1,1): 3,271 rows in total (≈3.39%)
- Two identical negatives (2,2,1 or 2,1,1,1): 20,777 rows in total (≈21.5%)
- All negatives unique (1,1,1,1,1): 71,436 rows (≈73.9%)
I realize the README already notes "Deduplified: No", but it might help future users if it also mentioned that multiple negative slots can carry identical text—especially that there are rows where all negatives are the same. When I first trained a model without checking for this, I unknowingly fed batches such as [pos1, neg1, same neg1, same neg1, same neg1…], and in my case that reduced downstream performance.
If it's feasible, calling this out explicitly in the README could make downstream cleaning or sampling decisions easier. Thanks again for maintaining the dataset, and I hope this feedback is helpful!