Balanced Toxicity Text Subset for Language Models

A curated subset of The Pile dataset focused on toxic text examples, balanced for training and evaluation. The dataset was created by researcher tomekkorbak and uploaded to Hugging Face in June 2022. It is part of a series of balanced subsets derived from the larger 825GB Pile corpus.

Use Cases

• Train a binary toxicity classifier using the 'text' field and inferred toxicity labels.
• Evaluate language model outputs for safety by comparing generated text against toxic examples in the dataset.
• Analyze linguistic patterns and markers associated with toxic language using the curated text samples.
• Balance training data for language models to reduce bias by incorporating this controlled subset.

Strengths

• Derived from The Pile, a large 825GB text corpus used for training state-of-the-art language models.
• Specifically balanced for toxicity, addressing class imbalance common in raw text data.
• Hosted on Hugging Face with available libraries like datasets, dask, and polars for efficient loading.

Limitations

• Exact row count and specific column structure are unknown from available metadata.
• Data is from 2022, which may not reflect the most current language or online toxicity trends.
• The balancing methodology and specific toxicity criteria are not detailed in the provided information.

Provenance

Source

Subset of The Pile dataset.

Collection Method

Curated and balanced selection by the author.

Freshness

Last updated in June 2022.

Geography

Region tag suggests 'us', but specific coverage is unknown.