Tag
New research suggests that with sufficient compute, filtering training data for language models may be unnecessary, and models can benefit from low-quality data.