princeton-nlp/QuRatedPajama-260B

QuRatedPajama

Paper: QuRating: Selecting High-Quality Data for Training Language Models

A 260B token subset of cerebras/SlimPajama-627B, annotated by princeton-nlp/QuRater-1.3B with sequence-level quality ratings across 4 criteria:

• Educational Value - e.g. the text includes clear explanations, step-by-step reasoning, or questions and answers
• Facts & Trivia - how much factual and trivia knowledge the text contains, where specific facts and obscure trivia are preferred over more common knowledge
• Writing Style - how polished and good is the writing style in the text
• Required Expertise: - how much required expertise and prerequisite knowledge is necessary to understand the text

In a pre-processing step, we split documents in into chunks of exactly 1024 tokens. We provide tokenization with the Llama-2 tokenizer in the input_ids column.

Guidance on Responsible Use:

In the paper, we document various types of bias that are present in the quality ratings (biases related to domains, topics, social roles, regions and languages - see Section 6 of the paper). Hence, be aware that data selection with QuRating could have unintended and harmful effects on the language model that is being trained. We strongly recommend a comprehensive evaluation of the language model for these and other types of bias, particularly before real-world deployment. We hope that releasing the data/models can facilitate future research aimed at uncovering and mitigating such biases. Note that the quality ratings do not measure the social or literary value of a text and should not be used for textual or demographic studies.

Citation:

@article{wettig2024qurating,
   title={QuRating: Selecting High-Quality Data for Training Language Models},
   author={Alexander Wettig, Aatmik Gupta, Saumya Malik, Danqi Chen},
   journal={arXiv preprint 2402.09739},
   year={2024}
}