MATHSCALE: SCALING INSTRUCTION TUNING FOR MATHEMATICAL REASONING

Summary from the TLDR AI newsletter

Researchers used synthetic data to generate 2 million path problems. They then trained a 7B model and found strong performance against state-of-the-art large language models.