README.md · tomekkorbak/pile-toxicity-balanced at main

Generation procedure

The dataset was constructed using documents from the Pile scored using using Perspective API toxicity scores.

The procedure was the following:

A chunk of the Pile (3%, 7m documents) was scored using the Perspective API.
The first half of this dataset is tomekkorbak/pile-toxic-chunk-0, 100k most toxic documents of the scored chunk
The first half of this dataset is tomekkorbak/pile-nontoxic-chunk-0, 100k least toxic documents of the scored chunk
Then, the dataset was shuffled and a 9:1 train-test split was done

The average scores of the good and bad half are 0.0014 and 0.67, respectively. The average score of the whole dataset is 0.33; the median is 0.51.

However, the weighted average score (weighted by document length) is 0.45. Correlation between score and document length is 0.2.

Score histogram:

Mean score per Pile subset