27 rows from test are also present in train
#2
by mbakshi1094 - opened
While working with this dataset, I found that 27 rows in the test split are also present in train. You can confirm with this query in the Data Studio SQL interface:
SELECT test.text_id, test.text, test.text_type, test.source, test.text = train.text
FROM test
JOIN train USING (text_id)
ORDER BY test.text_type, test.source, test.text_id;
26 of those rows (all the human_written ones) are also present in test_llama.
Thank you so much for flagging this! It looks like there was a bug with the hashing for text_id, so the texts were identical, but text_ids were different. I've now removed the duplicates and their AI-edited veresions from test_llama, as well as removed a couple of duplicate texts that were still hanging around.
katherinethai changed discussion status to closed