Toxicity-Balanced Text Corpus for Language Model Safety

Pile Toxicity Balanced2 is a text dataset designed for training and evaluating language models on toxic content. The dataset, created by researcher tomekkorbak, was uploaded to Hugging Face in April 2022. It is part of a series of datasets derived from The Pile, a large-scale text corpus used for AI development.

Use Cases

• Training binary classifiers to detect toxic vs. non-toxic text segments.
• Evaluating language model propensity to generate harmful completions by analyzing its text samples.
• Mitigating toxicity in model outputs by fine-tuning on this balanced subset of The Pile corpus.
• Studying the distribution of specific toxic language features (e.g., hate speech, threats) within a large text collection.

Strengths

• Derived from The Pile, a foundational corpus containing over 800 GB of diverse text data.
• Specifically balanced for toxicity, a critical property for reducing bias in safety evaluations.
• Hosted on Hugging Face with supporting libraries like `datasets` and `polars` for efficient access.

Limitations

• Specific row count, column structure, and exact balancing methodology are not documented.
• The underlying data from The Pile has a cutoff around 2020, limiting coverage of recent language and events.
• Potential label noise or subjective judgments inherent in toxicity classification tasks.

Provenance

Source

Subset of The Pile, a large-scale open-source language modeling dataset.

Collection Method

Curated and balanced from The Pile based on toxicity labels or scores.

Time Range

Reflects the temporal coverage of The Pile, primarily pre-2020.

Freshness

Last updated on the platform in April 2022; the source data is older.

Geography

Likely reflects the English-language and web-crawl origins of The Pile, with a bias towards US-centric sources.