Datasets:
Obvious low-quality data with high scores. Disappointing...
Duplicated low-quality template-generated data in ultrafineweb-en-part-0001-of-2048.parquet. Howerver, the score is 0.9933. Big problem.
Number of rows containing the string: 218
content score source
1389 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
1984 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
3008 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
3691 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
3855 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
... ... ... ...
549292 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
551020 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
552382 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
553207 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
554865 Enter captcha code\nWhy is this required? We c... 0.9933 FineWeb
{"source": "FineWeb", "text": "Enter captcha code\nWhy is this required? We can provide desired results with guaranteed correctness only to humans. Therefore, we need to verify that you are a human and not a computer. This website has been inundated with requests from remote servers instead of people, which caused a high load. By forcing this captcha on you, we can ensure that we offer you desired results.", "meta": {"score": 0.9933}}
Thanks for your reply.
Firstly, we did not use any dedup methods in our Ultra-FineWeb. Hence, it may get this case.
Also, fastText is a classifier that we cannot ensure all "high-quality" samples are correct, and we cannot recall all positive samples.
I believe this method can be optimized.