## Generation procedure The dataset was constructed using documents from [the Pile](https://pile.eleuther.ai/) scored using using [Perspective API](http://perspectiveapi.com) toxicity scores. The procedure was the following: 1. A chunk of the Pile (3%, 7m documents) was scored using the Perspective API. 1. The first half of this dataset is [tomekkorbak/pile-toxic-chunk-0](https://huggingface.co/datasets/tomekkorbak/pile-toxic-chunk-0), 100k *most* toxic documents of the scored chunk 2. The first half of this dataset is [tomekkorbak/pile-nontoxic-chunk-0](https://huggingface.co/datasets/tomekkorbak/pile-nontoxic-chunk-0), 100k *least* toxic documents of the scored chunk 3. Then, the dataset was shuffled and a 9:1 train-test split was done ## Basic stats 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: ![](https://huggingface.co/datasets/tomekkorbak/pile-toxicity-balanced/resolve/main/hist.png) Mean score per Pile subset | pile_set_name | score | length | |:------------------|----------:|------------:| | ArXiv | 0.141808 | 9963.82 | | Books3 | 0.405541 | 8911.67 | | DM Mathematics | 0.535474 | 8194 | | Enron Emails | 0.541136 | 1406.76 | | EuroParl | 0.373395 | 4984.36 | | FreeLaw | 0.279582 | 8986.73 | | Github | 0.495742 | 2184.86 | | Gutenberg (PG-19) | 0.583263 | 4034 | | HackerNews | 0.617917 | 3714.83 | | NIH ExPorter | 0.0376628 | 1278.83 | | OpenSubtitles | 0.674261 | 14881.1 | | OpenWebText2 | 0.613273 | 2634.41 | | PhilPapers | 0.549582 | 9693 | | Pile-CC | 0.525136 | 2925.7 | | PubMed Abstracts | 0.0388705 | 1282.29 | | PubMed Central | 0.235012 | 7418.34 | | StackExchange | 0.590904 | 2210.16 | | USPTO Backgrounds | 0.0100077 | 2086.39 | | Ubuntu IRC | 0.598423 | 4396.67 | | Wikipedia (en) | 0.0136901 | 1515.89 | | YoutubeSubtitles | 0.65201 | 4729.52 |